Alteration of the corpus callosum in patients with Alzheimer's disease: Deep learning-based assessment.

BACKGROUND: Several studies have reported changes in the corpus callosum (CC) in Alzheimer's disease. However, the involved region differed according to the study population and study group. Using deep learning technology, we ensured accurate analysis of the CC in Alzheimer's disease.
METHODS: We used the Open Access Series of Imaging Studies (OASIS) dataset to investigate changes in the CC. The individuals were divided into three groups using the Clinical Dementia Rating (CDR); 94 normal controls (NC) were not demented (NC group, CDR = 0), 56 individuals had very mild dementia (VMD group, CDR = 0.5), and 17 individuals were defined as having mild and moderate dementia (MD group, CDR = 1 or 2). Deep learning technology using a convolutional neural network organized in a U-net architecture was used to segment the CC in the midsagittal plane. Total CC length and regional magnetic resonance imaging (MRI) measurements of the CC were made.
RESULTS: The total CC length was negatively associated with cognitive function. (beta = -0.139, p = 0.022) Among MRI measurements of the CC, the height of the anterior third (beta = 0.038, p <0.0001) and width of the body (beta = 0.077, p = 0.001) and the height (beta = 0.065, p = 0.001) and area of the splenium (beta = 0.059, p = 0.027) were associated with cognitive function. To distinguish MD from NC and VMD, the receiver operating characteristic analyses of these MRI measurements showed areas under the curves of 0.65-0.74. (total CC length = 0.705, height of the anterior third = 0.735, width of the body = 0.714, height of the splenium = 0.703, area of the splenium = 0.649).
CONCLUSIONS: Among MRI measurements, total CC length, the height of the anterior third and width of the body, and the height and area of the splenium were associated with cognitive decline. They had fair diagnostic validity in distinguishing MD from NC and VMD.

Introduction

The corpus callosum (CC) is a wide and thick intermediate white matter track that includes several numbers of fibers that connect the two cortical hemispheres and provides interhemispheric transmission of information within the brain [1,2]. Most CC fibers have homotopic connectivity and topographical arrangement [3,4]. Shape changes or decreased CC size are associated with cognitive decline [5,6]. Two mechanisms of callosal atrophy have been proposed: Wallerian degeneration of axons in the white matter due to cortical cell death, and direct myelin breakdown and axonal damage of callosal fibers [7].

Many studies have reported callosal atrophy in patients with AD. Although AD is characterized by medial temporal atrophy, there are changes not only in the gray matter cortical area but also in the white matter tract. The CC is a structure composed only of white matter tract, we assumed that the CC would reflect the white matter change seen in AD. For this reason, CC was studies as a diagnostic marker for MCI and AD in previous studies. However, the regions involved in CC differ between studies. While some studies reported changes in the anterior part of the corpus callosum [8,9], other studies reported splenium atrophy [10,11]. This might be owing to the difference in the study population; however, it might also be due to the methods of segmenting the CC and the differences in defining the subregions [12].

Recently, advances in deep learning technology have made it possible to distinguish MRI images more accurately. Deep learning methods have an advantage in the segmentation of human brain structures for imaging analysis [13]. Traditional methods of imaging segmentation were intensity-based method, atlas-based method, or surface-based method. These methods were time-consuming and sensitive to noise [14]. Additionally, some traditional methods showed low accuracy due to anatomical variability, requiring handcrafting of the researcher for segmentation to increase the accuracy [15]. Using deep learning, it is possible to obtain more accurate results by performing segmentation using an automatic method. This increases the accuracy of medical imaging analysis. In deep learning methods, convolutional networks, which have been widely used in the past, are limited by the size of the available training sets and the size of the considered networks [16]. The U-net architecture is a method that performs very well in biomedical segmentation applications by supplementing the shortcomings of existing segmentation methods [17]. Therefore, in our study, we segmented the CC using deep learning methods, convolutional neural network organized in a U-net architecture, and then measured the characteristics of the CC, and based on this, we investigated the characteristics of the CC in Alzheimer’s disease.

Materials and methods

Patients

We used the Open Access Series of Imaging Studies (OASIS) cross-sectional dataset [18]. The OASIS dataset was freely released to the scientific community, consisting of 416 individuals aged 18–96 years, from a larger database of individuals who had participated in MRI studies at Washington University. Subjects with and without dementia were obtained from the longitudinal pool of the Washington University Alzheimer Disease Research Center (ADRC). Subjects with a primary cause of dementia other than AD, active neurological or psychiatric illness, serious head injury, history of clinically meaningful stroke, and use of psychoactive drugs were excluded. The determination of AD or control status is based solely on clinical methods. Clinical dementia rating (CDR) indicated dementia severity. CDR 0 indicated no dementia, CDR 0.5 indicated very mild dementia, and CDR 1 and 2 indicated mild and moderate dementia, respectively.

Of these, we excluded individuals aged less than 60 years to avoid the confounding effects of aging, which is known to influence CC size [19,20]. The individuals were divided into three groups. Individuals with CDR 0 were normal controls (NC), individuals with CDR 0.5 had very mild dementia (VMD), and individuals with CDR 1 and 2 had mild or moderate dementia (MD).

Image acquisition

For each participant, T1-weighted magnetization prepared rapid gradient-echo images were acquired using a 1.5-T vision scanner (Siemens, Erlangen, Germany) in a single imaging session with the following imaging parameters: sagittal slice thickness, 1.25 mm; no gap; repetition time, 9.7 ms; echo time, 4.0 ms; flip angle, 10°; inversion time, 20 ms; delay time, 200 ms; and resolution of 256 × 256 pixels (voxel size of 1 mm × 1 mm X 1.25mm).

Magnetic resonance imaging (MRI) processing

We used SPM12 for CC normalization and segmentation. All the human brain MRI volumes were normalized to the MNI 152 space using SPM12. MNI normalization in SPM12 brings all the individuals to the same space with a resolution of 79 × 95 × 79 and a voxel size of 2 × 2 × 2 mm. After normalization of all the individuals to the MNI space, segmentation was performed using the standard method of SPM12. SPM12 segments the MRI volumes into different parts, such as white matter (WM), gray matter, intracranial cerebral spinal fluid volume, bone, extracranial soft tissue, and background. However, for further preprocessing, we used only segmented WM. After the segmentation of MRI volumes, the midsagittal plane of the WM segmented volume was further used for extracting the CC region. The WM of the mid-sagittal plane was obtained by Talairach transformation using the alignment derived from the line connected the anterior commissure and the posterior commissure in SPM12. Then, segmented CC was passed to train the U-net model. The CC region was extracted by training with the U-net model. The extraction result was the voxel coordinates of the CC area. Using these coordinate values, each characteristic parameter was obtained with MATLAB. MATLAB-based code was developed to extract the CC region from the midsagittal planes of all individuals. Detailed MATLAB-based code was described in S1 Text.

The U-net consists of 23 convolutional layers in total in a contracting path, an expansive path, and a final layer. The contracting path consists of repeated applications of two 3 × 3 convolutions and a 2 × 2 max pooling operation with stride 2 for down-sampling. The expansive path consists of repeated applications of two 3 × 3 convolutions and a 2 × 2 convolution for up-sampling. At the final layer, a 1 × 1 convolution was used. It was trained with 16 batch sizes, 0.0001 learning rate for 200 epochs with early stopping, where the accuracy was obtained about 91.83%, where the training, validation, and test data set was splitted into a 70:10:20 ratio.

Definition of MRI measurements

The length of the CC can be measured easily by extracting the most inferior points of the rostrum and splenium. However, the distance measured by this method was assumed to be inaccurate. The purpose of this bisector line method is to calculate the length of the CC that includes its curvature [21]. In this method, the most inferior points (rows and columns) of the rostrum and splenium are identified. Then, the midpoint of the distance was assumed to be the center of the most inferior part of the rostrum and splenium. The midpoint served as the source for various numbers of radial chords that originated from the evaluated midpoint, each at a regular interval of 10°. This was performed using the interpolation method in MATLAB. Then, the length of the CC is measured using the Euclidean method for all the lines that connect the midpoints of these chords inside the CC (Fig 1A).

Fig 1

Estimation of MRI measurements (a) Evaluation of total CC length (b) Division of CC in five subregions (c) Measurements of each subregion (a, the width of the genu; b, the width of center between genu and rostrum; c, the width of rostrum; and d, height between the width of genu and rostrum; vertical lines of part 2, 3, 4; the height of the anterior, middle, and posterior body portion; lines of part 5, the rotatory diameter of region 5).

To perform the regional MRI measurements, first, using all the CC images, the extreme rows and columns were identified. (Length, L1 to L6) Then, using the estimated length (from L1 to L6), the callosal length of CC was divided into five equidistance subregions (Part 1 to Part 5) (Fig 1B). For the first fifth part, which corresponded to the genu; width of the genu (a), the width of the rostrum (c), center width of genu and rostrum (b), and lastly, height between the width of genu and width of the rostrum (d) were measured. For the second, third, and fourth positions that corresponded to the trunk, the height of the trunk was measured. This height was evaluated by measuring the height at the center of the center pixel for each region in the second, third, and fourth regions separately. The width of the body was evaluated by measuring the distance between L2 and L5. (Fig 1B) Then, the mean of all distances extracted in each region was detected. For the last fifth, which corresponded to the splenium, rotatory diameter measurement was performed. For this measurement, the fifth region (L5–L6) was extracted, and then the center was assumed to be the crossing point of the maximal horizontal diameter (L5–L6) and the maximal vertical diameter (mid column of horizontal diameter). Ten chords (diameter, D 1–D 10) were estimated using the interpolation method in MATLAB, the length of each starting from the maximal vertical diameter (= D1), at regular intervals of 10° until D 10. Then, using the Euclidean distance method, the length of each chord was estimated. Finally, the mean of all the extracted 10 chords, width, and height of the splenium was evaluated (Fig 1C).

Statistical analysis

The baseline characteristics were presented as mean values for continuous variables and percentages for categorical variables. Differences between each group were confirmed using analysis of variance (ANOVA) for continuous variables and chi-square tests for categorical variables. Differences in each MRI measurement among the groups were evaluated using ANOVA. The association of mini-mental state examination (MMSE) and each MRI measurement was evaluated using multiple linear regression analyses with MMSE as a determinant and each MRI measurement as an outcome variable after controlling for age, sex, education level, and estimated total intracranial volume (eTIV). Receiver operating curve (ROC) analyses of the MRI measurements were used to determine the optimal cutoff values with the Youden index. ROCs were compared among the measurements of the five MRI measurements in a pairwise manner. Statistical significance was defined as p<0.05. Statistical analyses were conducted using SPSS version 25 software (SPSS Inc., Chicago, IL, USA) and dBSTAT version 5 software.

Results

Demographics

The baseline demographics are presented in Table 1. The mean age of the patients in the MD group was the highest. The VMD group had the lowest number of women. There were no significant differences in education levels among the groups. (Table 1).

Table 1

Demographic and baseline characteristics.

	NC	VMD	MD	p-value
n	94	56	17
Age (mean±SD)	73.4±10.66	76.9±6.72	78.6±7.91	0.024
Sex (Female, %)	72 (76.6)	30 (53.6)	13 (76.5)	0.01
Education (less than college, %)	32 (34.0)	26 (46.4)	10 (58.8)	0.091
MMSE (mean±SD)	29.0±1.21	25.9±3.11	21.2±4.69	<0.0001
eTIV (mean±SD)	1447.24±150.962	1480.25±169.149	1446.65±101.174	0.422
MTA, right				<0.0001
0	39 (41.5)	9 (16.1)	0 (0)
1	38 (40.4)	29 (51.8)	5 (29.4)
2	15 (16.0)	14 (25.0)	6 (35.3)
3	2 (2.1)	4 (7.1)	6 (35.3)
4	0 (0)	0 (0)	0 (0)
MTA, left				<0.0001
0	41 (43.6)	5 (8.9)	0 (0)
1	34 (36.2)	27 (48.2)	2 (11.8)
2	16 (17.0)	15 (26.8)	8 (47.1)
3	3 (3.2)	8 (14.3)	6 (35.3)
4	0 (0)	1 (1.8)	1 (5.9)

n, number of individuals; NC, normal control; VMD, very mild dementia; MD, mild and moderate dementia; SD, standard deviation; MMSE, Mini-Mental State Examination; eTIV, estimated total intracranial volume; MTA, medial temporal lobe atrophy.

Comparisons of MRI measurements

There was no significant difference between individuals with normal cognition and individuals with cognitive impairment in all genu or rostrum-related measurements. A significant difference was observed in the height of the anterior third of the body between the NC and MD groups. The MD group also showed a shorter body width than that in the NC group. Both the VMD and MD groups demonstrated a shorter height of the splenium compared to the NC group. Both the VMD and MD groups showed less area of the splenium compared to the NC group. The MD group had a longer total CC length than that in the NC group (Table 2).

Table 2

Group differences of MRI measurements.

	NC	VMD	MD	p-value
Total CC length	47.40±2.818*	48.08±2.890	49.68±2.657*	0.008
Genu and rostrum
Width of the genu	4.20±1.141	4.00±1.335	3.53±1.328	0.105
Width of center between genu and rostrum	5.03±1.307	4.82±1.630	4.71±1.105	0.534
Width of rostrum	4.88±2.199	4.48±2.587	3.88±2.595	0.231
Height between the genu and rostrum	3.19±0.954	3.16±1.108	3.24±0.970	0.962
Body
Height of the anterior third	4.39±0.393*	4.27±0.436	4.05±0.571*	0.008
Height of the middle third	3.82±0.421	3.71±0.471	3.64±0.472	0.172
Height of the posterior third	3.69±0.524	3.52±0.514	3.47±0.469	0.077
Width	11.90±1.103*	11.50±1.067	11.16±1.327*	0.014
Splenium
Width	7.88±0.960	7.89±1.275	7.59±0.939	0.555
Height	5.48±0.936**	4.91±0.978*	4.59±0.712*	<0.0001
Area	6.27±1.313**	5.53±1.063*	5.46±1.135*	<0.0001

Data presented as means ± SD.

CC, corpus callosum; NC, normal control; VMD, very mild dementia; MD, mild and moderate dementia; SD, standard deviation.

Associations between MMSE and MRI measurements

Among parts of the CC, the body and splenium were associated with the MMSE score, while rostrum and genu did not correlate with the MMSE score. The height of the anterior and middle third of the body were associated with the MMSE score. Body width was also associated with the MMSE score. The height and area of the splenium were associated, while the width of the splenium was not associated with the MMSE score. Total CC length was inversely associated with the MMSE score (Table 3).

Table 3

Associations between MMSE score and each MRI measurements.

	Estimate	SE	R2	p-value
Total CC length	-0.139	0.06	0.242	0.022
Genu and rostrum
Width of the genu	0.026	0.026	0.223	0.325
Width of center between genu and rostrum	0.004	0.03	0.167	0.905
Width of rostrum	0.077	0.055	0.059	0.161
Height between the genu and rostrum	0.003	0.023	0.067	0.881
Body
Height of the anterior third	0.038	0.009	0.266	<0.0001
Height of the middle third	0.021	0.01	0.193	0.028
Height of the posterior third	0.017	0.011	0.142	0.137
Width	0.077	0.023	0.273	0.001
Splenium
Width	0.016	0.025	0.022	0.519
Height	0.065	0.019	0.364	0.001
Area	0.059	0.026	0.238	0.027

CC, corpus callosum; MMSE, Mini-Mental State Examination; MRI, magnetic resonance imaging; SE, standard error; R2, coefficient of determination.

When additional analysis was performed by correcting the diagnostic group, only the height of the anterior third of the body and the width of the body were associated with the MMSE score (S1 Table).

ROC curve analysis of CC measurements for discriminating dementia

ROC analysis of the MRI measurements of the CC to discriminate MD from NC and VMD revealed that the area under the curve (AUC) ranged from 0.65 to 0.74. At the cutoff values with the highest Youden indices, the width of the body showed high sensitivity, while the height of the splenium demonstrated the highest specificity. When the ROC analysis results were compared, there was no difference in the AUC among the MRI measurements (Table 4).

Table 4

Receiver operating curve analysis of corpus callosum measurements for distinguishing dementia from other groups.

	Total CC length	The height of the anterior third of the body	The width of the body	The height of the splenium	The area of the splenium
AUC	0.705	0.735	0.714	0.703	0.649
SE	0.059	0.073	0.063	0.059	0.071
95% CI	0.590–0.820	0.592–0.877	0.591–0.836	0.587–0.819	0.511–0.788
Sensitivity	0.706	0.706	0.824	0.529	0.588
Specificity	0.660	0.767	0.673	0.773	0.713
Cutoff value	48.760	4	11.250	4	5.330

CC, corpus callosum; AUC, area under the curve; SE, standard error; CI, confidence interval.

Discussion

Total CC length was negatively associated with cognitive function. Among the parts of the CC, the body and splenium were associated with cognitive function. Among the MRI measurements of the body, the height of the anterior third and total length of the body reflected cognitive function. Among the measurements of the splenium, height and area were more reflective of cognitive function. These MRI measurements were suitable for distinguishing MD from NC and VMD.

In this study, there were differences in the body and splenium, while there were no differences in MRI measurements between groups for rostrum and genu. The data registered in the OASIS dataset used in this study were obtained from patients diagnosed with AD [18]. Many previous studies have mainly investigated the association between CC atrophy and cognitive impairment in patients with AD [11,22,23]. In this study, both the anterior body part and splenium were affected, which was consistent with previous studies in which AD affected both the anterior and posterior parts of the CC [23-25]. When analyzing the association between MMSE and MRI measurements of the CC as well as comparison by disease group, the body, and splenium of the CC were also associated with MMSE. This study showed that the body and splenium could be markers reflecting the degree of cognitive function, regardless of the disease group. This is in line with a previous study that reported that CC atrophy was associated with global cognitive function even in normal elderly or elderly individuals with only mildly impaired cognitive function [26].

In particular, the size of the splenium in the VMD and MD groups was smaller than that in the normal control group, indicating that the splenium might be a more vulnerable region than other regions of the CC. A previous study also reported that CC atrophy already occurred in patients with mild cognitive impairment, and this study was an extension of the results [12]. Although in some previous studies, atrophy occurred from the anterior to the posterior direction [7,27] other studies reported that the callosal size decreased from the splenium [10,23]. Fibers of splenium connected with parietal, temporal, and occipital cortical regions [28]. One of the possible mechanisms of callosal atrophy is Wallerian degeneration, in which callosal fibers are lost due to the distal loss of the callosal projecting neurons; that is, CC atrophy might be a reflection of cortical neuronal loss [7]. AD is a disease that mainly causes temporal atrophy [29]. In patients with AD, the splenium was more likely to reflect atrophy than the rostrum.

In contrast, in the present study, the total CC length was the longest in the MD group. This was contrary to previous studies and our previous findings that the MD group showed more atrophy. This might indicate that the total CC length reflected less atrophy. In studies analyzing the shape of the CC in other diseases, regional CC thickness reflected the characteristics of the disease better than the total CC length among the measurements representing shape [30,31]. Similarly in our study, the height of the body and splenium reflected cognitive impairment better than the total length. Another possible explanation for why the total CC length was the longest in MD was that the total CC length might reflect distortion. In previous studies, CC circularity was found to be reduced in patients with AD [5]. The total CC length may reflect reduced circularity. A reduction in circularity could result from a deformation of the CC owing to disease-related enlargement of the lateral ventricles, reflecting an overall atrophic process. In this study, since the length of the CC of the mid-sagittal section was measured, not the volume of the whole CC, irregular changes in the CC due to the deformity caused by the whole-brain atrophy might have increased the total CC length.

Measurements of the body and splenium exhibited fair diagnostic validity for the discrimination of dementia. In particular, the height of the anterior third of the body showed the highest AUC. We also analyzed the value of measurements for discriminating VMD; however, it showed poor diagnostic validity. As for diagnostic methods, our measurements were more meaningful in distinguishing MD, particularly the height of the anterior third of the body, which appears to be useful. Through this study, it was found that changes in the CC were of greater diagnostic value in discriminating patients with mild AD rather than discriminating early in cognitive decline. In previous studies using diffusion tensor imaging, white matter integrity measurements of the CC were reported to have good diagnostic validity for the early detection of MCI and AD [32,33]. However, in previous studies that analyzed CC atrophy, atrophy was evident in AD but not in MCI [24,25]. We used only the morphology of the CC, which might explain why the diagnostic value in the VMD group was not clear, while the diagnostic value in the MD group was more pronounced.

In our study, for accurate measurements of the CC, we used deep learning, U-net, for automatic CC segmentation and extraction. The model was trained using the same features extracted from the developed method as input. The U-net architecture has a particularly good performance in segmentation applications [17]. Therefore, in this study, the accuracy of measurements of the CC might be improved. Although some of the previous research results were different from ours [8,10], we believe that the accuracy of our study would be higher than that of previous studies as we used the deep learning method to perform CC segmentation.

However, this study had several limitations. First, the study individuals were diagnosed clinically, and amyloid pathology was not considered for diagnosis. Second, since this was a cross-sectional study, changes in longitudinal CC atrophy were not considered. Third, since the study population, particularly the MD group, was smaller than the normal group, the possibility that the difference in number between these groups could influence the study results exists. Finally, we used only mid-sagittal plane cross-sectional images for analysis; therefore, we could not measure the total volume. Despite these limitations, we analyzed the characteristics of the CC using a more accurate and automated method using the deep learning method, which enabled us to identify various characteristics of the CC of AD. We anticipate providing new clinical insights into the association between the characteristics of the CC and AD.

Associations between MMSE score and each MRI measurements after controlling diagnostic group.

(DOCX)

Click here for additional data file.

The MATLAB-based code for extracting the corpus callosum.

(DOCX)

Click here for additional data file.

21 May 2021

PONE-D-21-14181

Alteration of the corpus callosum in patients with Alzheimer’s disease: Deep learning-based assessment

PLOS ONE

Dear Dr. Kim,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

Upon my own reading of the manuscript, like Reviewer 1, I had difficulty in understanding exactly how the deep learning method was developed and applied. As this is a key part of the manuscript, many more details need to be provided. Please see the recommendations of Reviewer 1 in this regard. Please also follow the suggestions of Reviewer 2 with respect to assessing the robustness/specificity of your findings.

Please submit your revised manuscript by Jul 05 2021 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

Please include the following items when submitting your revised manuscript:

A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

We look forward to receiving your revised manuscript.

Kind regards,

Niels Bergsland

Academic Editor

PLOS ONE

Journal Requirements:

When submitting your revision, we need you to address these additional requirements.

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

2. We note that you have stated that you will provide repository information for your data at acceptance. Should your manuscript be accepted for publication, we will hold it until you provide the relevant accession numbers or DOIs necessary to access your data. If you wish to make changes to your Data Availability statement, please describe these changes in your cover letter and we will update your Data Availability statement to reflect the information you provide.

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Partly

Reviewer #2: Yes

**********

2. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: Yes

**********

3. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: Yes

**********

4. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

**********

5. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: I want to thank the authors for presenting an interesting, and scientifically relevant, manuscript on measuring various aspects of the corpus callosum in patients with Alzheimer’s Disease (VMD and MD) versus normal controls. They present results showing how various features of the corpus callosum (such as total length, subsection height, width, area, and so on) relate to cognitive function. They also perform an ROC analysis for distinguishing the patient populations using the most relevant corpus callosum features. They have a particularly good explanation in the discussion section for why CC length may be increased due to enlargement of ventricles, thus distorting the CC.

Overall, the paper is well-written, with only minor English errors. The results are interesting, but there lacks an explanation of how they performed their deep learning algorithm, which, from my understanding, lays the foundation for their results. This algorithm is not mentioned in any meaningful way in the methods section. This makes the evaluation of the manuscript as a whole difficult.

Major points:

- Introduction:

o 3rd Paragraph on deep learning

� This paragraph needs to become clearer as it is too generalizing at the moment. What methods are being compared to the deep learning methods? There are both plenty of conventional programming methods (FreeSurfer, SPM, FSL, and so on), as well as many different types of deep learning methods (CNN, vector-based, PCA, and so on). There’s no need to explain all these in-depth, but it would improve understanding for the reader to know which ones you are referring to, without having to dive deep into the references given.

� Not all traditional methods require handcrafting of the researcher (for example, Freesurfer will "recon-all" without user intervention). Although a manual intervention may sometimes improve results.

- Materials and Methods:

o MRI processing section

� Did you use SPM12 for CC segmentation? Thus not a deep learning algorithm? It is not clear to me when the deep learning algorithm was applied.

� Tell us more about the MATLAB-based code that was used for extracting the CC. This is the main methodology of your research paper, if I understand correctly?

� How was the midsagittal slice decided? Or did you extract several slices in the mid-sagittal plane? The manner by which a midsagittal slice is extracted can influence the results quite greatly, as various angles will produce different results for the same CC.

- Discussion

o 6th Paragraph

� Here you bring up that you used a Unet for automatic segmentation and extraction. And if I understand correctly, it was trained on the segmentations from SPM12. It is however unclear what the actual accuracy of your Unet algorithm is. Did you perform a cross-validation? This paragraph should be in the methods section, and explain in-depth how you constructed your Unet. How many convolutions? Which optimizer? How many images were used for training and testing? How was it validated? Learning rate? Batch size?

Minor points:

- Abstract:

o Methods section

� Please add the abbreviation “OASIS” as this is a well-known dataset.

� Please include what type of deep learning technology you used, i.e.: “Deep learning using a convolutional neural network organized in a Unet fashion…”

o Results section

� Sometimes you write “MR measurements” and sometimes “MRI measurements”. Both are fine, but I think picking one or the other would improve the flow.

- Introduction:

o 4th Paragraph

� Again, it would be nice to know which deep learning method you are using.

� If using a Unet, there could be value in adding Ronneberger et al.’s article where the Unet was first presented.

- Materals and Methods:

o Patients section

� In the second paragraph you mention that you exclude individuals aged less than 60 years to avoid the confounding of aging, which is known to influence CC size. There is no reference for this. Here’s a reference saying that there is no size difference due to age (which is what makes CC such a promising biomarker for patients with diseases where it actually does atrophy). PMID: 11445261.

o Image acquisition section

� A voxel is a three-dimensional pixel. If I understand your dataset correctly, it should state: "1 mm X 1 mm X 1.25 mm".

o MRI processing section

� What are you referring to as "soft tissue"? Aren't both white matter and gray matter soft tissues?

- Discussion

o 1st Paragraph

� Mixing of past tense and present tense in two adjacent sentences, which I believe are both referring to the results of your paper (i.e. they should both be in past tense).

o Last Paragraph

� How was the mid-sagittal slice decided upon and extracted?

Reviewer #2: The present study describes characteristics of the corpus callosum in older individuals with normal cognition (CDR=0), ‘mild dementia’ (CDR=0.5), and dementia (CDR=1-2), assesses associations between corpus callosum characteristics and MMSE score, and the diagnostic discriminative value of corpus callosum characteristics. I do not have the proper background to review the deep learning methods used in the paper, but I have some remarks based on other aspects of the paper:

- It seems that the terms ‘dementia’ and ‘Alzheimer’s disease (AD)’ are used interchangeably throughout the paper. However, not all patients with dementia have AD. Some additional information on the diagnostic background of the cases would be helpful to clarify this. Was the clinical diagnosis for all cases with ‘mild dementia’ and ‘dementia’ Alzheimer’s disease? What diagnostic criteria were used? How was ‘normal cognition’ assessed?

- When analyzing the associations between corpus callosum characteristics and MMSE score, it would be helpful to control for diagnosis (as a dummy variable) to make sure that the associations are not driven by diagnostic groups.

- The relevance of assessing corpus callosum characteristics in Alzheimer’s disease is unclear. Why are the authors interested in this structure? What could be the added value of this marker compared with for example medial temporal lobe atrophy (i.e., what could be the added diagnostic discriminative value of corpus callosum characteristics on top of established measures such as medial temporal lobe atrophy)?

- Similar to the previous remark: The authors find associations between corpus callosum characteristics and cognition, but is this not a reflection of cerebral atrophy in general? Could the authors control for cortical/global atrophy in their analyses? Or perhaps white matter atrophy?

- Is any additional information available on the MRI characteristics of these cases, such as vascular damage / medial temporal lobe atrophy scores? It would be helpful to include this for better characterization of the study sample.

**********

6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: Yes: Michael Platten

Reviewer #2: Yes: Whitney Freeze

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

7 Aug 2021

Manuscript: PONE-D-21-14181

Title: Alteration of the corpus callosum in patients with Alzheimer’s disease: Deep learning-based assessment

Reviewer #1: I want to thank the authors for presenting an interesting, and scientifically relevant, manuscript on measuring various aspects of the corpus callosum in patients with Alzheimer’s Disease (VMD and MD) versus normal controls. They present results showing how various features of the corpus callosum (such as total length, subsection height, width, area, and so on) relate to cognitive function. They also perform an ROC analysis for distinguishing the patient populations using the most relevant corpus callosum features. They have a particularly good explanation in the discussion section for why CC length may be increased due to enlargement of ventricles, thus distorting the CC.

Overall, the paper is well-written, with only minor English errors. The results are interesting, but there lacks an explanation of how they performed their deep learning algorithm, which, from my understanding, lays the foundation for their results. This algorithm is not mentioned in any meaningful way in the methods section. This makes the evaluation of the manuscript as a whole difficult.

Major points:

1.Introduction:

o 3rd Paragraph on deep learning

1-1. This paragraph needs to become clearer as it is too generalizing at the moment. What methods are being compared to the deep learning methods? There are both plenty of conventional programming methods (FreeSurfer, SPM, FSL, and so on), as well as many different types of deep learning methods (CNN, vector-based, PCA, and so on). There’s no need to explain all these in-depth, but it would improve understanding for the reader to know which ones you are referring to, without having to dive deep into the references given.

In response to reviewer’s comment, we added sentences to the third paragraph of introduction section as follows. (Introduction section, Page 5, Line 20 – Page 6, Line 7)

“Traditional methods of imaging segmentation were intensity-based method, atlas-based method, or surface-based method. These methods were time-consuming and sensitive to noise. [14] Additionally, some traditional methods showed low accuracy due to anatomical variability, requiring handcrafting of the researcher for segmentation to increase the accuracy. [15] Using deep learning, it is possible to obtain more accurate results by performing segmentation using an automatic method. This increases the accuracy of medical imaging analysis. In deep learning methods, convolutional networks, which have been widely used in the past, are limited by the size of the available training sets and the size of the considered networks. [16] The u-net architecture is a method that performs very well in biomedical segmentation applications by supplementing the shortcomings of existing segmentation methods. [17]”

1-2. Not all traditional methods require handcrafting of the researcher (for example, Freesurfer will "recon-all" without user intervention). Although a manual intervention may sometimes improve results.

In response to reviewer’s comment, we modified the third paragraph of introduction section as follows. (Introduction section, Page 5, Line 22 – Page 6, Line 2)

“Additionally, some traditional methods showed low accuracy due to anatomical variability, requiring handcrafting of the researcher for segmentation to increase the accuracy. [15] Using deep learning, it is possible to obtain more accurate results by performing segmentation using an automatic method.”

2. Materials and Methods:

o MRI processing section

2-1. Did you use SPM12 for CC segmentation? Thus not a deep learning algorithm? It is not clear to me when the deep learning algorithm was applied.

Response: We used SPM 12 for CC segmentation. Then, segmented CC was passed to train the U-net model. The CC region was extracted by training with U-net model. This content was added to the Materials and Methods section as follow. (Materials and Methods section, Page 8, Line 16 - Line 18)

“Then, segmented CC was passed to train the U-net model. The CC region was extracted by training with the U-net model. The extraction result was the voxel coordinates of the CC area.”

2-2. Tell us more about the MATLAB-based code that was used for extracting the CC. This is the main methodology of your research paper, if I understand correctly?

Response: The MATLAB code for extracting the CC was as follows.

a=spm_vol('c2orig0001.nii'); % real the volume

b=spm_read_vols(a); % read the volume into image matrix

data=squeeze(b(128,:,:)); % extract the sagittal slice; since data is 79x95x79, we will take 39th/40th slice

figure(1),imshow(data,[]); % show the image (extract_midslice.jpg)

bw=bwareaopen(data,100); % remove small objects; objects with pixels less than 100

labelbw=bwlabel(bw); % label the image

figure(3),imshow(bw);% display labeled image (labeled_image.jpg)

figure(4),imshow(labelbw==2); % extract the appropriate label image which has corpus callosum (Segmented_CC.jpg)

We added this code as supplementary text.

2-3. How was the midsagittal slice decided? Or did you extract several slices in the mid-sagittal plane? The manner by which a midsagittal slice is extracted can influence the results quite greatly, as various angles will produce different results for the same CC.

In response to reviewer’s comment, we added the method of mid-sagittal plane decision in Materials and Methods section as follows. (Materials and Methods section, Page 8, Line 22 - Page 9, Line 7)

“The WM of the mid-sagittal plane was obtained by Talairach transformation using the alignment derived from the line connected the anterior commissure and the posterior commissure in SPM12.”

3. Discussion

o 6th Paragraph

Here you bring up that you used a Unet for automatic segmentation and extraction. And if I understand correctly, it was trained on the segmentations from SPM12. It is however unclear what the actual accuracy of your Unet algorithm is. Did you perform a cross-validation? This paragraph should be in the methods section, and explain in-depth how you constructed your Unet. How many convolutions? Which optimizer? How many images were used for training and testing? How was it validated? Learning rate? Batch size?

In response to reviewer’s comment, we added sentences to the Materials and Methods section as follows. (Materials and Methods section, Page 8, Line 22 - Page 9, Line 7)

“The u-net consists of 23 convolutional layers in total in a contracting path, an expansive path, and a final layer. The contracting path consists of repeated applications of two 3 × 3 convolutions and a 2 × 2 max pooling operation with stride 2 for down-sampling. The expansive path consists of repeated applications of two 3 × 3 convolutions and a 2 × 2 convolution for up-sampling. At the final layer, a 1 × 1 convolution was used. It was trained with 16 batch sizes, 0.0001 learning rate for 200 epochs with early stopping, where the accuracy was obtained about 91.83%, where the training, validation, and test data set was splitted into a 70:10:20 ratio.”

Minor points:

1. Abstract:

1-1. Methods section

1-1-1. Please add the abbreviation “OASIS” as this is a well-known dataset.

Response: We added the abbreviation “OASIS”. (Abstract section, Page 3, Line 7)

1-1-2. Please include what type of deep learning technology you used, i.e.: “Deep learning using a convolutional neural network organized in a Unet fashion…”

Response: As your recommendation, we modified this sentence as follows. (Abstract section, Page 3, Line 11 – Line 13)

“Deep learning technology using a convolutional neural network organized in a U-net architecture was used to segment the CC in the midsagittal plane.”

1-2. Results section

Sometimes you write “MR measurements” and sometimes “MRI measurements”. Both are fine, but I think picking one or the other would improve the flow.

Response: In response to reviewer’s comment, we changed MR measurements to MRI measurements.

2. Introduction:

2-1. 4th Paragraph

2-1-1. Again, it would be nice to know which deep learning method you are using.

Response: In response to reviewer’s comment, we modified the fourth paragraph of introduction section as follows. (Introduction section, Page 6, Line 7 – Line 10)

“Therefore, in our study, we segmented the CC using deep learning methods, convolutional neural network organized in a U-net architecture, and then measured the characteristics of the CC, and based on this, we investigated the characteristics of the CC in Alzheimer's disease.”

2-1-2. If using a Unet, there could be value in adding Ronneberger et al.’s article where the Unet was first presented.

Response: In response to reviewer’s comment, we added the Ronneberger et al.’s article as reference number 17.

17. Ronneberger O, Fischer P, Brox T, editors. U-Net: Convolutional Networks for Biomedical Image Segmentation2015; Cham: Springer International Publishing.

2-2. Materals and Methods:

2-2-1. Patients section

In the second paragraph you mention that you exclude individuals aged less than 60 years to avoid the confounding of aging, which is known to influence CC size. There is no reference for this. Here’s a reference saying that there is no size difference due to age (which is what makes CC such a promising biomarker for patients with diseases where it actually does atrophy). PMID: 11445261.

Response: Thank you for your recommendation. However, we found that there were studies that reported corpus callosum size decline in aging. Therefore, we added these studies as references.

22. Hopper KD, Patel S, Cann TS, Wilcox T, Schaeffer JM. The relationship of age, gender, handedness, and sidedness to the size of the corpus callosum. Academic Radiology. 1994;1(3):243-8. doi: 10.1016/s1076-6332(05)80723-8.

23. Janowsky JS, Kaye JA, Carper RA. Atrophy of the corpus callosum in Alzheimer's disease versus healthy aging. J Am Geriatr Soc. 1996;44(7):798-803. Epub 1996/07/01. doi: 10.1111/j.1532-5415.1996.tb03736.x. PubMed PMID: 8675927.

2-2-2. Image acquisition section

A voxel is a three-dimensional pixel. If I understand your dataset correctly, it should state: "1 mm X 1 mm X 1.25 mm".

Response: In response to reviewer’s comment, this issue has been corrected as recommended by reviewer. (Materials and Methods section, Page 8, Line 3)

2-2-3. MRI processing section

What are you referring to as "soft tissue"? Aren't both white matter and gray matter soft tissues?

In this article, soft tissue meant extracranial soft tissue, and in response to the reviewer’s comment, we changed ‘soft tissue’ to ‘extracranial soft tissue’. (Materials and Methods section, Page 8, Line 12)

3. Discussion

3-1. 1st Paragraph

� Mixing of past tense and present tense in two adjacent sentences, which I believe are both referring to the results of your paper (i.e. they should both be in past tense).

In response to reviewer’s comment, we changed the sentences to the past tense.

3-2. Last Paragraph

� How was the mid-sagittal slice decided upon and extracted?

In response to reviewer’s comment, we added the method of mid-sagittal plane decision in Materials and Methods section as follows. (Materials and Methods section, Page 8, Line 22 - Page 9, Line 7)

“The WM of the mid-sagittal plane was obtained by Talairach transformation using the alignment derived from the line connected the anterior commissure and the posterior commissure in SPM12.”

Reviewer #2: The present study describes characteristics of the corpus callosum in older individuals with normal cognition (CDR=0), ‘mild dementia’ (CDR=0.5), and dementia (CDR=1-2), assesses associations between corpus callosum characteristics and MMSE score, and the diagnostic discriminative value of corpus callosum characteristics. I do not have the proper background to review the deep learning methods used in the paper, but I have some remarks based on other aspects of the paper:

1. It seems that the terms ‘dementia’ and ‘Alzheimer’s disease (AD)’ are used interchangeably throughout the paper. However, not all patients with dementia have AD. Some additional information on the diagnostic background of the cases would be helpful to clarify this. Was the clinical diagnosis for all cases with ‘mild dementia’ and ‘dementia’ Alzheimer’s disease? What diagnostic criteria were used? How was ‘normal cognition’ assessed?

Response: This study used the OASIS dataset, and the definitions of subjects revealed in the OASIS dataset are as follows.

“Subjects aged 18 to 96 years were selected from a larger database of individuals who had participated in MRI studies at Washington University. Especially, older subjects, aged 60 and older, with and without dementia were obtained from the longitudinal pool of the Washington University Alzheimer Disease Research Center (ADRC). Older adults underwent the ADRC’s full clinical assessment as described below. Subjects with a primary cause of dementia other than AD, active neurological or psychiatric illness, serious head injury, history of clinically meaningful stroke, and use of psychoactive drugs were excluded, as were subjects with gross anatomical abnormalities evident in their MRI images. Dementia status was established and staged using the CDR scale. The determination of AD or control status is based solely on clinical methods, without reference to psychometric performance, and any potential alternative causes of dementia must be absent. The diagnosis of AD is based on clinical information that the subject has experienced gradual onset and progression of decline in memory and other cognitive and functional domains. A global CDR of 0 indicates no dementia, and CDRs of 0.5, 1, 2, and 3 represent very mild, mild, moderate, and severe dementia respectively. These methods allow for the clinical diagnosis of AD in individuals with a CDR of 0.5 or greater, based on standard criteria, that is confirmed by histopathological examination in 93% of the individuals, even for those in the earliest symptomatic stage (CDR 0.5) of AD who elsewhere may be considered to represent ‘mild cognitive impairment”.

Therefore, dementia in our study refers to Alzheimer’s disease.

In response to reviewer’s comment, we added sentences in Patients of Materials and Methods section as follow. (Materials and Methods section, Page 7, Line 6 - Line 11)

“Subjects with and without dementia were obtained from the longitudinal pool of the Washington University Alzheimer Disease Research Center (ADRC). Subjects with a primary cause of dementia other than AD, active neurological or psychiatric illness, serious head injury, history of clinically meaningful stroke, and use of psychoactive drugs were excluded. The determination of AD or control status is based solely on clinical methods.”

2. When analyzing the associations between corpus callosum characteristics and MMSE score, it would be helpful to control for diagnosis (as a dummy variable) to make sure that the associations are not driven by diagnostic groups.

Response: As your recommendation, we further controlled for the diagnostic group in the analysis of the associations between corpus callosum characteristics and MMSE score as below. We added this results as Supplementary table e-1.

Supplementary table e-1. Associations between MMSE score and each MRI measurements after controlling diagnostic group

Estimate SE R2 p-value

Total CC length -0.037 0.082 0.260 0.655

Genu and rostrum

Width of the genu 0.002 0.036 0.229 0.945

Width of center between genu and rostrum 0.001 0.042 0.168 0.977

Width of rostrum 0.060 0.076 0.061 0.427

Height between the genu and rostrum 0.001 0.032 0.069 0.983

Body

Height of the anterior third 0.046 0.012 0.273 <0.0001

Height of the middle third 0.022 0.013 0.193 0.103

Height of the posterior third 0.015 0.016 0.143 0.331

Width 0.083 0.032 0.273 0.010

Splenium

Width 0.009 0.035 0.030 0.799

Height 0.031 0.025 0.381 0.222

Area 0.016 0.036 0.263 0.656

And we also these results added to Results section as follows. (Results section, Page 15, Line 4 - Line 6)

“When additional analysis was performed by correcting the diagnostic group, only the height of the anterior third of the body and the width of the body were associated with the MMSE score. (Supplementary table e-1)”

3. The relevance of assessing corpus callosum characteristics in Alzheimer’s disease is unclear. Why are the authors interested in this structure? What could be the added value of this marker compared with for example medial temporal lobe atrophy (i.e., what could be the added diagnostic discriminative value of corpus callosum characteristics on top of established measures such as medial temporal lobe atrophy)?

Response: Although Alzheimer’s disease is characterized by medial temporal atrophy, it has been known in previous studies that changes in gray matter cortical area as well as white matter tracts are seen. The corpus callosum is a structure composed only of white matter tract, and we investigated the characteristics of the corpus callosum, it would reflect the white matter change seen in AD. We added these to the introduction section as follows. (Introduction section, Page 5, Line 9 - Line 13)

“Although AD is characterized by medial temporal atrophy, there are changes not only in the gray matter cortical area but also in the white matter tract. The CC is a structure composed only of white matter tract, we assumed that the CC would reflect the white matter change seen in AD. For this reason, CC was studies as a diagnostic marker for MCI and AD in previous studies.”

4. Similar to the previous remark: The authors find associations between corpus callosum characteristics and cognition, but is this not a reflection of cerebral atrophy in general? Could the authors control for cortical/global atrophy in their analyses? Or perhaps white matter atrophy?

Response: Our findings are already estimated total intracranial volume (eTIV)-corrected values. Therefore, we think it represents the change of the corpus callosum that is not affected by global atrophy.

5. Is any additional information available on the MRI characteristics of these cases, such as vascular damage / medial temporal lobe atrophy scores? It would be helpful to include this for better characterization of the study sample.

Response: In response to reviewer’s comment, we added the medial temporal lobe atrophy score to Table 1.

Table 1. Demographic and baseline characteristics

NC VMD MD p-value

n 94 56 17

Age (mean±SD) 73.4±10.66 76.9±6.72 78.6±7.91 0.024

Sex (Female, %) 72 (76.6) 30 (53.6) 13 (76.5) 0.01

Education (less than college, %) 32 (34.0) 26 (46.4) 10 (58.8) 0.091

MMSE (mean±SD) 29.0±1.21 25.9±3.11 21.2±4.69 <0.0001

eTIV (mean±SD) 1447.24±150.962 1480.25±169.149 1446.65±101.174 0.422

MTA, right <0.0001

0 39 (41.5) 9 (16.1) 0 (0)

1 38 (40.4) 29 (51.8) 5 (29.4)

2 15 (16.0) 14 (25.0) 6 (35.3)

3 2 (2.1) 4 (7.1) 6 (35.3)

4 0 (0) 0 (0) 0 (0)

MTA, left <0.0001

0 41 (43.6) 5 (8.9) 0 (0)

1 34 (36.2) 27 (48.2) 2 (11.8)

2 16 (17.0) 15 (26.8) 8 (47.1)

3 3 (3.2) 8 (14.3) 6 (35.3)

4 0 (0) 1 (1.8) 1 (5.9)

Submitted filename: Reviewers comments.docx

Click here for additional data file.

23 Aug 2021

PONE-D-21-14181R1

Alteration of the corpus callosum in patients with Alzheimer’s disease: Deep learning-based assessment

PLOS ONE

Dear Dr. Kim,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

Please carefully consider the point regarding cross-validation, as suggested by the Reviewer.

Please submit your revised manuscript by Oct 07 2021 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

Please include the following items when submitting your revised manuscript:

A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

We look forward to receiving your revised manuscript.

Kind regards,

Niels Bergsland

Academic Editor

PLOS ONE

Journal Requirements:

Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice.

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #1: All comments have been addressed

Reviewer #2: All comments have been addressed

**********

2. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Partly

Reviewer #2: Yes

**********

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: Yes

**********

4. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: Yes

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

**********

6. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: Thanks to the authors for their re-working of the manuscript. You have done a good job.

There is just one last thing that I think is necessary before acceptance. You have now presented the "training, validation, and testing" numbers. You have an accuracy of 91% with early termination.

- Firstly, it's unclear if this accuracy is reflecting the validation or testing. I assume testing, but it could be both, as it seemed to have influenced an "early stop" when, I again assume, you reached the highest accuracy level.

There's a bias in exiting training early at the highest accuracy level, as it can represent an overfitting of your own data.

Considering the above stated, in combination with the fact that you only have a small sample size (100 - which is not atypical for biomedical scenarios - and thus why you aptly chose the U-net). I would like to see a cross-validation of your data (I'd recommend at least K-fold: 10, unless you want to do leave-one-out cross-validation). This will give you a better and more true reflection of the performance of your algorithm. It is not uncommon to just do a cross-validation with training and validation data (i.e. skip the test data -> you need as much as you can for the training). Also, have a pre-set number of epochs, and choose the accuracy of the last epoch (i.e. not necessarily the "highest accuracy"). When you do the cross-validation there should not be any tuning of the hyperparameters, as this will affect the results. In the end we are interested in how your algorithm performs without having to tweak it every single time it's applied.

Also, please describe your metric "accuracy". I believe a dice-score metric would be the absolute best for your scenario, but several papers present the classic accuracy (TP + TN / TP + TN + FP + FN), as this tends to show higher accuracy.

Reviewer #2: The authors have addressed all of the comments sufficiently. I have no further comments on the manuscript.

**********

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: Yes: Michael Platten

Reviewer #2: Yes: Whitney Freeze

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

7 Oct 2021

In response to reviewer’s comment, we performed 10-fold cross-validation as follows. And we used Intersection-Over_Union (IOU) Matric for accuracy (TP/TP+FP+FN)

Plots of accuracies and losses of validation per each k

k Accuracies and losses of validation data Accuracies of test data

1 0.974044

2 0.974546

3 0.974053

4 0.974123

5 0.973244

6 0.974320

7 0.974123

8 0.974245

9 0.974196

10 0.973881

Submitted filename: Reviewer_s comments_2_results.docx

Click here for additional data file.

12 Oct 2021

Alteration of the corpus callosum in patients with Alzheimer’s disease: Deep learning-based assessment

PONE-D-21-14181R2

Dear Dr. Kim,

We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements.

Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication.

An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at http://www.editorialmanager.com/pone/, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to-date. If you have any billing related questions, please contact our Author Billing department directly at authorbilling@plos.org.

If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

Kind regards,

Niels Bergsland

Academic Editor

PLOS ONE

Additional Editor Comments (optional):

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #1: All comments have been addressed

**********

2. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

**********

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

**********

4. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

**********

6. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: Thank you for answering the comments. No further comments or questions from me. I wish you good luck.

**********

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

14 Dec 2021

PONE-D-21-14181R2

Alteration of the corpus callosum in patients with Alzheimer’s disease: Deep learning-based assessment

Dear Dr. Kim:

I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department.

If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact onepress@plos.org.

If we can help with anything else, please email us at plosone@plos.org.

Thank you for submitting your work to PLOS ONE and supporting open access.

Kind regards,

PLOS ONE Editorial Office Staff

on behalf of

Dr. Niels Bergsland

Academic Editor

PLOS ONE