Whole lesion histogram analysis of apparent diffusion coefficient predicts therapy response in locally advanced rectal cancer.

BACKGROUND: Whole-tumor apparent diffusion coefficient (ADC) histogram analysis is relevant to predicting the neoadjuvant chemoradiation therapy (nCRT) response in patients with locally advanced rectal cancer (LARC). AIM: To evaluate the performance of ADC histogram-derived parameters for predicting the outcomes of patients with LARC.
METHODS: This is a single-center, retrospective study, which included 48 patients with LARC. All patients underwent a pre-treatment magnetic resonance imaging (MRI) scan for primary tumor staging and a second restaging MRI for response evaluation. The sample was distributed as follows: 18 responder patients (R) and 30 non-responders (non-R). Eight parameters derived from the whole-lesion histogram analysis (ADCmean, skewness, kurtosis, and ADC10th, 25th, 50th, 75th, 90th percentiles), as well as the ADCmean from the hot spot region of interest (ROI), were calculated for each patient before and after treatment. Then all data were compared between R and non-R using the Mann-Whitney U test. Two measures of diagnostic accuracy were applied: the receiver operating characteristic curve and the diagnostic odds ratio (DOR). We also reported intra- and interobserver variability by calculating the intraclass correlation coefficient (ICC).
RESULTS: Post-nCRT kurtosis, as well as post-nCRT skewness, were significantly lower in R than in non-R (both P < 0.001, respectively). We also found that, after treatment, R had a larger loss of both kurtosis and skewness than non-R (∆%kurtosis and ∆skewness, P < 0.001). Other parameters that demonstrated changes between groups were post-nCRT ADC10th, ∆%ADC10th, ∆%ADCmean, and ROI ∆%ADCmean. However, the best diagnostic performance was achieved by ∆%kurtosis at a threshold of 11.85% (Area under the receiver operating characteristic curve [AUC] = 0.991, DOR = 376), followed by post-nCRT kurtosis = 0.78 × 10-3 mm2/s (AUC = 0.985, DOR = 375.3), ∆skewness = 0.16 (AUC = 0.885, DOR = 192.2) and post-nCRT skewness = 1.59 × 10-3 mm2/s (AUC = 0.815, DOR = 168.6). Finally, intraclass correlation coefficient analysis showed excellent intraobserver and interobserver agreement, ensuring the implementation of histogram analysis into routine clinical practice.
CONCLUSION: Whole-tumor ADC histogram parameters, particularly kurtosis and skewness, are relevant biomarkers for predicting the nCRT response in LARC. Both parameters appear to be more reliable than ADCmean from one-slice ROI. ©The Author(s) 2022. Published by Baishideng Publishing Group Inc. All rights reserved.

Core Tip: Whole-tumor apparent diffusion coefficient (ADC) histogram analysis is an emergent imaging analysis in which every voxel is used to obtain a histogram; it thus provides statistical information about tumors. Our study revealed that ADC histogram profiling is a valuable approach that can help differentiate treatment response in locally advanced rectal cancer. When determining tailored treatments that are associated with minimal morbidities, such as the watch and wait method, an accurate treatment response prediction is critical. Given the limitations of this study, more research is needed to establish the clinical utility of our findings.

INTRODUCTION

Neoadjuvant chemoradiation therapy (nCRT) is the gold standard treatment for patients with locally advanced rectal cancer (LARC), followed by surgical resection and adjuvant chemotherapy[1,2]. After nCRT, the ability to achieve tumor reduction or even a pathological complete response (pCR) is observed in approximately 75% of treated patients, whereas the remainder exhibited no treatment response[3,4]. The ability to predict the response to nCRT is important for patients with potentially curable LARC who wish to explore personalized treatment to expand their therapeutic outcomes[5].

Functional magnetic resonance imaging (MRI) techniques, such as diffusion-weighted imaging (DWI), can provide additional physiological information about a tumor’s cellular environment, offering great potential to evaluate the therapeutic response to nCRT[5]. This is because the apparent diffusion coefficient (ADC), a quantitative parameter used to assess water diffusion through tissue in DWI, shows an inverse relationship with tissue cellularity[6]. Viable tumor cells restrict the mobility of water, whereas necrotic tumor cells allow the increased diffusion of water molecules[7].

The possibility that ADC may be associated with the nCRT response has been amply investigated in LARC[8-12]; however, significant correlations have not been found in any studies to date[10]. Inconsistencies in previous findings may be due to a lack of standardized imaging and acquisition techniques[5,11], but they may also be due to the fact that the ADC measurements were performed using a manually drawn region of interest (ROI) from a single slice of the ADC map, which holds limited ability to reflect the actual whole-tumor characteristics[13-15].

In the case of whole-lesion histogram analysis of the ADC, a volumetric ROI is positioned on the entire lesion over contiguous slices and a histogram of ADC values reflecting voxel frequency is constructed, leading to the improved evaluation of heterogeneity[16]. Based on this method, first-order heterogeneity parameters can be obtained, which assess the spectrum of ADC values gained from all voxels within a volume of interest[17]. A growing number of studies have used ADC histogram parameters, as these analyses provide additional information that can aid in the discrimination between benign and malignant regions, or they can help to better characterize the response to treatment in different tumors, such as ovarian, prostate, and breast cancer[18-21]. The application of whole-volume ADC histogram analysis in rectal tumors is increasing in frequency as well, and the role of this parameter in predicting nCRT is promising but limited[22-25].

The purpose of this study was to investigate the imaging response to nCRT using DWI in patients with LARC. We hypothesized that the ADC histogram-derived parameter might better predict treatment responses to nCRT compared with ADC from the hotspot ROI, as histogram parameters can display the heterogeneous features of tumors.

MATERIALS AND METHODS

Patients

The institutional review board approved this retrospective study, and the requirement to obtain informed consent was waived given the study’s retrospective nature. The study population was selected from LARC patients at our institution between February 2015 and October 2020. According to Enkhbaatar et al[23], we defined the inclusion criteria as follows: (1) Proven histopathology of rectal adenocarcinoma; (2) greater than stage T2 on pre-nCRT MR imaging; with or without regional lymph node metastases and no distant metastases; (3) pre- and post-nCRT rectal MRI imaging with diffusion-weighted (DW) imaging; (4) long-course nCRT; and (5) surgical resection. Mucinous tumors were excluded from this study.

Forty-eight patients were enrolled in the study (34 men and 14 women; age range: 28–84 years). All patients were further divided into two subgroups based on the pathological response of the primary tumor: responders (R) and non-responders (non-R). Only patients with grade 0 according to the TRG-Ryan system were regarded as patients with a complete pathological response (R), while patients with TRG 1-3 were non-R.

MRI protocol

All images were obtained on a 3T MRI system (Discovery MR 750w GEM®; General Electric Healthcare, Milwaukee, WI, United States) using a phased-array body coil. Intravenous antispasmodic agents were not administered, and patients received no bowel preparation before the MRI examination. Our study groups comprised patients who underwent pre-treatment MRI for primary tumor staging, and a second restaging MRI for response evaluation 6 wk after the completion of nCRT. The scanning protocol is listed in Table 1[23]. In brief, we obtained standard T2-weighted (T2W) spin-echo sequences in axial, coronal, and sagittal directions. To improve tumor tissue visualization (including the delineation of the muscular layer), these planes were planned perpendicular to the main axis of the tumor. Moreover, a T1W spin-echo sequence in an axial direction, as well as an axial non-enhanced DWI with b = 1200 s/mm2, were acquired. ADC maps were automatically generated using the in-line software provided by the vendor during image acquisition. Additionally, axial, sagittal, and coronal fat-suppressed contrast T1W sequences were acquired and used to suppress the signal from adipose tissue. A gadolinium-based contrast agent (Gd-DTPA, Magnevist; Bayer Schering, Berlin, Germany) was used to enhance the quality of MRI. Representative images of our MRI protocol are provided in Figure 1.

Table 1

Magnetic resonance imaging sequences and data acquisition parameters

Parameter	Magnetic resonance imaging sequences
	T2 FSE sagittal	T2 FSE axial	T2 FSE coronal	T1 FSE axial	DWI axial	T1 + GD axial	T1 + GD coronal	T1 + GD coronal
Repetition time in ms	5325	9890	7509	850	7750	435	295	265
Echo time in ms	102	102	102	Min	Min	Min	Min	Min
Slices, n	30	40	30	40	40	40	40	30
FOV	24	20	20	20	20	20	20	24
Slices thickness in mm	4	4	4	4	4	4	4	4
Broadband in Hz/Px	62.5	62.5	50	62.5	-	50	50	50
Phase	384	384	416	384	60	320	320	320
Acquisition time in min:s	2:35	3:08	2:45	3:53	5:18	2:31	2:16	2:02

DWI: Diffusion-weighted imaging; FSE: Fast spin-echo; GD: Gadolinium; MRI: Magnetic resonance imaging.

Figure 1

Representative images of magnetic resonance imaging protocol. A-C: Axial T2 (A), apparent diffusion coefficient (ADC) map and T2 fusion ADC map color (B) and images of bulky tumor (C), showing tumor extending more than 5 mm into the mesorectal fat and invading the mesorectal fascia.

Image analysis

Two radiologists (JARP and MEJ, with 10 years and 5 years of experience in gastrointestinal imaging, respectively) reviewed the imaging studies and performed all tumor measurements on the pre- and post-nCRT images. At the initial review, each radiologist was blinded to the other radiologist’s opinion. Also, they were blinded to the pathology results to assess interobserver and intraobserver variability. After that, the two radiologists would hold a discussion to arrive at a final decision by consensus. If a disagreement occurred, another radiologist with 25 years of experience (YVN) aided in making the final decision.

DWI analysis was performed with a workstation using the GE Advantage Workstation 4.6 software featuring the READYVIWER application (2006–2010; General Electric, Boston, MA, United States). On the pre-nCRT b1200 diffusion images, the tumor was defined as a focal mass with high signal intensity in comparison with the signal of the normal adjacent rectal wall. More precisely, the delineated ROIs covered the edge of each lesion, and the ROIs were drawn along the inner margin of the rectal walls to avoid intraluminal gas, water, and other contents. Further, necrotic areas, cysts, and vessels related to each lesion at the corresponding slice were also avoided, as identified on T2WI images. In addition, the highest and lowest slices of the DWI images were excluded given their partial volume effects[24]. After nCRT, the tumor was defined by focal areas of residual high signal, as identified on the b1200 images within the location of the primary tumor bed and/or corresponding with the residual tumor on T2WI MRI images as a reference standard. To compare and identify the tumor location, the pre-treatment images were at the readers’ disposal when analyzing the post-treatment images.

It should be noted that, in the first instance, one large ROI was placed to cover most of the largest axial tumor cross-section, which facilitated the calculation of the ADCmean values (ROI ADCmean). Thereafter, a volume of interest (VOI) was manually created on the ADC maps, where ROIs were drawn on all tumor slices (whole-lesion measurement). Within this VOI, the following parameters were calculated: (1) ADCmean, the average ADC value of all voxels within the VOI; (2) ADCn% (10th, 25th, 50th, 75th, and 90th percentiles), the point at which the n% of the voxel values that formed the histogram were found to be at the left; (3) skewness, which measures the asymmetry of the distribution of values about the mean value; and (4) kurtosis, which is a measure of the ‘peakedness’ of the distribution of values in the ROI image. The corresponding frequency table for each lesion was exported, and the histogram parameters were computed by SPSS v. 26.0 (IBM Corporation, Armonk, NY, United States). Figure 2 is a schematic illustration of a representative ROI.

Figure 2

Images of rectal tumor before and after neoadjuvant chemoradiation therapy. A, C: T2-weighted magnetic resonance images obtained in 67-year-old man with a rectal tumor (histopathologic response Ryan 1) to evaluate tumor volume; B, D: Diffusion-weighted images (DWI) that were obtained from the same case. As we can see in the present case, regions of interest were drawn manually slice by slice on DWI images along the edge of the lesion to cover as much tumor area as possible without excluding cystic or necrotic areas. nCRT: Neoadjuvant chemoradiation therapy.

Histopathologic review

Specimens were evaluated according to an established protocol that was previously described by our research team[26]. In brief, fresh surgical specimens were evaluated to determine the quality of the mesorectal excision before being fixed in 4% formaldehyde for 48 h prior to sectioning. After fixation, the specimens were serially sectioned (in slices of 1 cm), and the mesorectal boundary was linked. When the residual tumor was visible, a minimum submission of four blocks was recommended. All mesorectal lymph nodes were histologically examined, as was the involvement of the circumferential resection margin. When no residual tumor cells were identified, each block was cut into 3 level sections, and immuhistochemistry for keratin was done. All hematoxylin and eosin slides were reviewed by an experienced pathologist (EHB, with 15 years of experience examining rectal cancer).

The pathologic response of the primary tumor was estimated using the modified Ryan’s classification as follow[26,27]: TRG0, complete response with no viable cancer cells; TRG1 moderate response with single cancer cells or small groups of cancer cells; TRG2, minimal response with residual cancer outgrown by fibrosis, and TRG3, poor response with minimal or no tumor killing and extensive residual cancer.

Statistical analyses

The following formula was used to calculate changes in all metrics included in the current study: PerC = (Parameter post-treatment - Parameter pre-treatment) / Parameter pre−treatment × 100.

It must be noted that when pre- and post-nCRT kurtosis values were obtained, a result of +3.00 indicated the absence of kurtosis. To simplify the interpretation, we adjusted this result to 0 (i.e. kurtosis of –3 = 0). Thus, any reading other than 0 was referred to as an excess of kurtosis. On the other side, to negate division by 0 when calculating the percentage change in kurtosis, we added 3, i.e. [(Kurtosis post-treatment + 3) – (Kurtosis pre-treatment + 3) / (Kurtosis pre-treatment + 3)] × 100.

In the case of skewness, and to avoid dividing by 0, only change (not the percentage change) was used (i.e. skewness post treatment – skewness pre-treatment)[28]. As skewness did not have a lower bound such as kurtosis, the +/– sign was considered to calculate changes in this parameter. To compare variables among R and non-R, a Mann–Whitney U test (MWU) was applied, as the Kolmogorov-Smirnov test confirmed the non-normal distribution of any parameter included here. Accordingly, the data were presented as medians and interquartile ranges (IQR)[29]. When the differences in a variable were significant (P < 0.05) in the MWU test, the cut-off value, sensitivity, specificity, positive predictive value, negative predictive value, area under the receiver operating characteristic (ROC) curve (AUC), and accuracy, were analyzed. The optimal cut-off values of ADCmean from the hot spot ROI and parameters derived from the histogram analysis of DWI were determined via the Youden index, while differences in the AUC were analyzed according to the method described by DeLong et al[30]. Furthermore, the diagnostic odd ratio (DOR) was designed to provide an additional measure of the performance of our potentially useful biomarkers to predict treatment response in LARC.

Finally, the intraobserver variability and interobserver variability were assessed using the intraclass correlation coefficient (ICC). For the agreement analysis, the outcomes were interpreted as follows, in accordance with Cicchetti (1994): 0.2 or less, poor agreement; 0.21–0.40, fair agreement; 0.41–0.60, moderate agreement; 0.61–0.74, good agreement; and 0.75–1.00, excellent agreement[31]. Statistical analyses were performed using SPSS v. 26. P < 0.05 was considered statistically significant.

RESULTS

Among the 58 patients that were originally included in this study, 10 had severe imaging artifacts. Thus, our final sample included 48 patients whose clinical and pathological characteristics are described in Table 2.

Table 2

Clinical and pathological characteristics of the patients’ studies

Characteristics	n (%)
Sex
Female	23 (48)
Male	25 (52)
RECIST 1.1
Partial response	22 (46)
Stable disease	13 (27)
Progressive disease	13 (27)
Ryan’s classification
0	18 (38)
1	10 (21)
2	9 (19)
3	11 (22)
Treatment response
Complete responders after nCRT	18 (38)
Non-responders’ patients after nCRT	30 (62)
Tumor location
Upper third	7 (15)
Middle third	14 (29)
Lower third	20 (41)
Diffuse	7 (15)
ypT stage
T0	4 (8)
T1s	4 (8)
T1a	2 (4)
T2	8 (17)
T3	20 (42)
T4b	10 (21)
ypN stage
N0	20 (42)
N1a	14 (29)
Nic	14 (29)
Degree of differentiation
Well-differentiated adenocarcinoma	6 (13)
Moderately differentiated adenocarcinoma	35 (73)
Poorly differentiated adenocarcinoma	7 (14)
Surgical approach
Low anterior resection	16 (33)
Intersphincteric resection	26 (54)
Abdominoperineal resection	6 (13)

nCRT: Neoadjuvant chemoradiation therapy.

The median values and IQRs for the ROI ADCmean values and parameters derived from the histogram analysis of DWI are described in Table 3. Accordingly, post-nCRT kurtosis, as well as post-nCRT skewness, were significantly lower in R than in non-R (both P < 0.001, respectively). Furthermore, our results showed significant differences in the relative changes of kurtosis (∆%kurtosis) between R and non-R (P < 0.001), with the largest loss of kurtosis in R. Additionally, median ∆skewness displayed lower values in R than in non-R (P < 0.001).

Table 3

Median and interquartile range of pre- and post-neoadjuvant chemoradiation therapy parameters, as well as of changes between pre- and post-treatment values

	Responders	Non-responders	P value
Pre-nCRT parameters
pre-nCRT ADCmean	0.75 (0.60-0.90)	0.85 (0.70-0.90)	0.146
10th percentile	0.20 (0.17-0.23)	0.20 (0.17-0.26)	0.812
25th percentile	0.32 (0.29-0.35)	0.32 (0.23-0.41)	1.000
50th percentile	0.40 (0.35-0.40)	0.40 (0.30-0.50)	0.698
75th percentile	0.56 (0.47-0.58)	0.57 (0.45-0.63)	0.391
90th percentile	0.71 (0.55-0.77)	0.76 (0.50-0.80)	0.556
Skewness	1.10 (0.90-1.14)	1.19 (0.88-1.37)	0.135
Kurtosis	0.89 (0.83-0.92)	0.92 (0.83-0.95)	0.296
ROI ADCmean	0.92 (0.80-1.20)	0.91 (0.83-0.93)	0.562
Post- nCRT parameters
post- nCRT ADCmean	1.20 (0.98-1.52)	1.10 (0.90-1.30)	0.065
10th percentile	0.36 (0.30–0.37)	0.32 (0.31–0.34)	0.0361
25th percentile	0.41 (0.40-0.52)	0.42 (0.41-0.50)	0.476
50th percentile	0.66 (0.56-0.70)	0.65 (0.51-0.66)	0.127
75th percentile	0.71 (0.67-0.80)	0.70 (0.66-0.75)	0.050
90th percentile	0.89 (0.80-0.95)	0.80 (0.79-0.89)	0.105
Skewness	0.92 (0.60–1.14)	2.00 (1.15–2.67)	< 0.0011
Kurtosis	0.65 (0.59–0.72)	0.90 (0.80–0.90)	< 0.0011
ROI ADCmean	2.50 (1.50-2.70)	2.00 (1.80-2.30)	0.056
Changes between pre-treatment and post-treatment
∆%ADCmean	57% (14%–103%)	27% (0%–59%)	0.0321
∆%ADC10th	86% (37%–118%)	48% (11%–88%)	0.0201
∆%ADC25th	39% (19%-57%)	22% (0%-58%)	0.905
∆%ADC50th	70% (31%-86%)	31% (2%-65%)	0.067
∆%ADC75th	40% (20%-61%)	37% (0%-57%)	0.288
∆%ADC90th	27% (11%-58%)	9% (7%-119%)	0.061
∆skewness	–0.20 (–0.40–0.00)	0.49 (0.10–0.50)	< 0.0011
∆%kurtosis	41% (18%–54%)	2.5% (1.4%–5.9%)	< 0.0011
ROI ∆%ADCmean	55% (48%–60%)	23% (15%–30%)	0.0201

Statistically significant difference. ADC: Apparent diffusion coefficient; nCRT: Neoadjuvant chemoradiation therapy; ROI: Region of interest.

We also found that patients with a favorable response (R) had higher post-nCRT ADC10th values than did non-R (P = 0.036). Correspondingly, the median values of ∆%ADC10th, ∆%ADCmean, and ROI ∆%ADCmean were also higher in R than in non-R, (P = 0.020, P = 0.032 and P = 0.020, respectively).

Receiver operating characteristics of those parameters that exhibited significant differences in the MWU test are reported in Table 4. The highest AUC values for predicting the treatment response in LARC were demonstrated by ∆%kurtosis, post-nCRT kurtosis, ∆skewness and post-nCRT skewness (AUCs = 0.991, 0.985, 0.885, and 0.815, respectively). Meanwhile, the lowest diagnostic accuracy was observed in post-nCRT ADC10th (AUC = 0.681), ∆%ADCmean (AUC = 0.686), ∆%ADC10th (AUC = 0.589) and ROI ∆%ADCmean (AUC = 0.583).

Table 4

Diagnostic performance of the best magnetic resonance imaging histogram derived parameters to detect responder patients

	Cutoff	Sensitivity	Specificity	PPV	NPV	Accuracy	AUC (95%CI)
∆%kurtosis	11.85%	94.4%	96.7%	94.4%	96.7%	96.0%	0.991 (0.925-1.000)
Post-nCRT kurtosis	0.78	93.3%	99.0%	90%	99.0%	96.0%	0.985 (0.957-1.000)
∆skewness	0.16	66.7%	99.0%	64.3%	99.0%	79.2%	0.885 (0.795-0.975)
Post-nCRT skewness	1.59	63.3%	99.0%	62.0%	99.0%	77.1%	0.815 (0.795-0.634)
Post-nCRT ADC10th	0.34×10-3 mm2/s	66.7%	73.3%	60.0%	79.0%	71.0%	0.681 (0.509-0.852)
∆%ADCmean	56.00%	56.0%	77.0%	56.0%	73.3%	66.7%	0.686 (0.500-0.820
∆%ADC10th	74.21%	61.1%	70.0%	55.0%	75.0%	66.7%	0.589 (0.483-0.815)
ROI ∆%ADCmean	55.00%	61.0%	69.0%	52.0%	72.0%	65.3%	0.583 (0.425-0.715)

ADC: Apparent diffusion coefficient; AUC: Area under the receiver operating characteristic curve; nCRT: Neoadjuvant chemoradiation therapy; NPV: Negative predictive value; PPV: Positive predictive value; ROI: Region of interest.

The ROC curves for ∆%kurtosis, post-nCRT kurtosis, ∆skewness, and post-nCRT skewness are displayed in Figure 3, while the comparison of AUC values between all of our potentially useful biomarkers for predicting the treatment response in LARC are presented in Supplementary Table 1.

Figure 3

Receiver operating characteristic curves displaying the diagnostic performances of the four histogram parameters derived from apparent diffusion coefficient values with the highest accuracy. A: ∆%kurtosis; B: Post-neoadjuvant chemoradiation therapy (nCRT) kurtosis; C: ∆skewness; D: Post-nCRT skewness. AUC: Area under the receiver operating characteristic curve.

It is important to mention that according to the DeLong analysis, no significant differences were found in the diagnostic accuracy of ∆%kurtosis and post-nCRT kurtosis. As well, no differences were demonstrated between ∆skewness and post-nCRT skewness. However, the latter two parameters had lower accuracy than kurtosis-derivate metrics.

Finally, to verify the diagnostic accuracy of all metrics reported in Table 4, we calculated DORs. The DOR of a test is the ratio of the odds of positivity if a patient has a disease relative to the odds of positivity when a patient does not have a disease. The value of DOR ranges from 0 to infinity, with higher values indicating better discriminatory test performance[32,33]. As demonstrated in Table 5, ∆%kurtosis and post-nCRT kurtosis had the highest power of discrimination for treatment response by using DORs (approximately 376), followed by ∆skewness (192.2) and post-nCRT skewness (168.6). Meanwhile, the lowest power of discrimination was observed in post-CRT ADC10th (5.48), ∆%ADCmean (4.26), ∆%ADC10th (3.65), and ROI ∆%ADCmean (3.47).

Table 5

Diagnostic odds ratios of magnetic resonance imaging parameters in differentiating respond and non- respond patients in locally advanced rectal cancer

	Diagnostic odds ratio	95%CI
∆%kurtosis	376.0	228.9-842.1
Post-nCRT kurtosis	375.3	225.7-887.7
∆skewness	192.2	69.0-253.3
Post-nCRT skewness	168.6	54.0-251.7
Post-nCRT ADC10th	5.48	1.0-19.6
∆%ADCmean	4.26	1.0-14.4
∆%ADC10th	3.65	1.0-12.5
ROI ∆%ADCmean	3.47	1.0-11.2

ADC: Apparent diffusion coefficient; nCRT: Neoadjuvant chemoradiation therapy; ROI: Region of interest.

Regarding interobserver and intraobserver variability, the parameters derived from the histogram analysis of DWI, as well as the ADC values from the hotspot ROI, had an excellent agreement. The ICC measuring intraobserver variability ranges from 0.777-0.931 (Table 6), while the ICC measuring intraobserver variability ranges from 0.889-0.993 (Table 7).

Table 6

Intraobserver variability

		ICC	95%CI
Basal
Test1 and test2, reader1	ROI ADCmean	0.850	0.742–0.800
Test1 and test2, reader2	ROI ADCmean	0.890	0.850-0.820
After treatment
Test1 and test2, reader1	ROI ADCmean	0.800	0.750-0.819
Test1 and test2, reader2	ROI ADCmean	0.823	0.800-0.850
Basal
Test1 and test2, reader1	ADCmean	0.850	0.756-0.920
Test1 and test2, reader2	ADCmean	0.777	0.745-0.812
After treatment
Test1 and test2, reader1	ADCmean	0.845	0.830-0.850
Test1 and test2, reader2	ADCmean	0.823	0.800-0.833
Basal
Test1 and test2, reader1	10th percentile	0.820	0.880-0.950
Test1 and test2, reader2	10th percentile	0.880	0.800-0.920
After treatment
Test1 and test2, reader1	10th percentile	0.780	0.740-0.853
Test1 and test2, reader2	10th percentile	0.853	0.723-0.901
Basal
Test1 and test2, reader1	25th percentile	0.803	0.800-0.922
Test1 and test2, reader2	25th percentile	0.863	0.801-0.895
After treatment
Test1 and test2, reader1	25th percentile	0.788	0.750-0.837
Test1 and test2, reader2	25th percentile	0.820	0.780-0.846
Basal
Test1 and test2, reader1	50th percentile	0.850	0.840-0.920
Test1 and test2, reader2	50th percentile	0.845	0.790-0.860
After treatment
Test1 and test2, reader1	50th percentile	0.821	0.800-0.913
Test1 and test2, reader2	50th percentile	0.833	0.800-0.897
Basal
Test1 and test2, reader1	75th percentile	0.821	0.790-0.860
Test1 and test2, reader2	75th percentile	0.859	0.820-0.920
After treatment
Test1 and test2, reader1	75th percentile	0.851	0.790-0.880
Test1 and test2, reader2	75th percentile	0.837	0.791-0.856
Basal
Test1 and test2, reader1	90th percentile	0.850	0.820-0.890
Test1 and test2, reader2	90th percentile	0.880	0.850-0.960
After treatment
Test1 and test2, reader1	90th percentile	0.831	0.800-0.902
Test1 and test2, reader2	90th percentile	0.901	0.850-0.975
Basal
Test1 and test2, reader1	Skewness	0.920	0.900-0.940
Test1 and test2, reader2	Skewness	0.901	0.880-0.923
After treatment
Test1 and test2, reader1	Skewness	0.931	0.920-0.950
Test1 and test2, reader2	Skewness	0.889	0.877-0.910
Basal
Test1 and test2, reader1	Kurtosis	0.920	0.890-0.950
Test1 and test2, reader2	Kurtosis	0.910	0.850-0.960
After treatment
Test1 and test2, reader1	Kurtosis	0.890	0.850-0.960
Test1 and test2, reader2	Kurtosis	0.880	0.840-0.982

ADC: Apparent diffusion coefficient; ICC: Intraclass correlation coefficient; ROI: Region of interest.

Table 7

Interobserver variability (intraclass correlation coefficient and 95% confidence intervals)

Pre-treatment	Reader one vs reader two	Post-treatment	Reader one vs reader two
ROI ADCmean	0.985 (1.900-0.999)	ROI ADCmean	0.889 (0.850-0.950)
ADCmean	0.989 (0.980–0.994)	ADCmean	0.990 (0.985-0.995)
10th percentile	0.972 (0.951–0.984)	10th percentile	0.992 (0.986–0.996)
25th percentile	0.970 (0.947–0.983)	25th percentile	0.950 (0.940-0.982)
50th percentile	0.986 (0.976–0.992)	50th percentile	0.987 (0.945-0.995)
75th percentile	0.989 (0.980–0.994)	75th percentile	0.990 (0.982–0.994)
90th percentile	0.989 (0.980–0.994)	90th percentile	0.972 (0.987–0.996)
Skewness	0.990 (0.982–0.994)	Skewness	0.993 (0.987–0.996)
Kurtosis	0.992 (0.986–0.995)	Kurtosis	0.972 (0.951–0.984)

ADC: Apparent diffusion coefficient; ROI: Region of interest.

DISCUSSION

Heterogeneity of malignant lesions is a feature that can be determined by characterizing changes in the histogram analysis of ADC values, which is recognized as a promising tool in cancer research when discerning between benign and malignant tumors or to better characterize the response to anti-cancer treatments[34-38].

This study focused on the ADCmean from the hot-spot ROI and a series of parameters corresponding to certain points on the ADC histogram using DWI, which have been proposed to predict treatment response in patients with rectal cancer[39,40]. As our results demonstrated, the parameters that changed significantly in response to nCRT were ∆%kurtosis, post-nCRT kurtosis, ∆skewness, post-nCRT skewness, post-nCRT ADC10th, ∆%ADCmean, ∆%ADC10th, and ROI ∆%ADCmean. However, the highest diagnostic accuracy was obtained for ∆%kurtosis, post-nCRT kurtosis, post-nCRT skewness, and ∆skewness, suggesting that these metrics might be useful when selecting responders (TRG 0) for an organ preservation approach with either ‘watch-and-wait’ or local excision[39,40].

The results derivate from parameters with the highest diagnostic accuracy in predicting treatment response to nCRT in the current work are reviewed below.

First, we demonstrated that both post-nCRT kurtosis and post-nCRT skewness were significantly lower in R than in non-R. The overall trends from the histogram studies have shown that, following treatment, the histogram analysis of DWI and diffusion kurtosis imaging (DKI) shifted to the right upon decreased kurtosis and skewness in rectal cancer[39-43]. For example, in 2017, Hu et al[39] reported that the post-treatment mean kurtosis derived from DKI showed reduced values in R when compared with non-R patients, whereas Enkhbaatar et al[23] (2019) documented that the histogram of R presented negative changes in skewness following a loss of this parameter after therapy.

In the same way, kurtosis from R had greater reductions than from non-R, which indicates Gaussian or flatter distributions in patients with a complete response to the therapy. In biological tissues, it is believed that the non-Gaussian behavior (more precisely, a platykurtic curve) of water might occur because of a heterogeneous environment characterized by multiple compartments, organelles, and semipermeable membranes[44]. Thus, when an important reduction in kurtosis is noticed, a higher displacement of water molecules in DWI is assumed.

Furthermore, as mentioned above, negative changes of skewness after nCRT were seen in R, while non-R exhibited positive changes in this parameter. Negatively skewed curves show the majority of scores above the mean, and positively skewed curves are just the opposite[44]. In physiology, the association between changes in skewness and responses to antineoplastic therapy have not been fleshed out, but a curve negatively skewed suggests a loss of cellular structure[23]. Therefore, favorable treatment response is suspected.

Our MWU analysis also found differences between R and non-R across other parameters, such as ADC10th, ∆%ADCth, ∆%ADCmean and ROI∆%ADCmean, as stated in our results section. However, both the ROC curve analysis and the DOR calculation indicated that only ∆%kurtosis, post-nCRT kurtosis, ∆skewness, and post-nCRT skewness appear to predict a favorable response to the therapy, whereas the other metrics did not possess that predictive property.

Briefly, the Youden index calculation indicated that post-nCRT kurtosis, post-nCRT skewness, and ∆skewness values below 0.78 × 10-3 mm2/s, 1.59 × 10-3 mm2/s, and 0.16, respectively, might be significant indicators of the occurrence of pCR. Meanwhile, ∆% changes above 11.85% also indicated a positive treatment effect with high accuracy. It is important to remember that, according to the DeLong analysis, the kurtosis-related parameters exhibit a better diagnostic performance than do skewness-related parameters.

Aligned with this finding, numerous authors have documented that kurtosis is more directly correlated to the underlying structural, physiological, molecular, and metabolic changes that occur during tumor progression than skewness[45]. This may be the reason why the kurtosis of ADC values has been used to indicate deviations from Gaussianity, even in the most challenging mathematical designs that predict the response to chemotherapy, such as radiomics analysis[46-48].

The results obtained from the ROC curved are partially supported by the estimated DORs, which were approximately 376 for both ∆%kurtosis and post-nCRT kurtosis. This means that for the cut-off points of ∆%kurtosis and post-nCRT kurtosis calculated here, the odds for positivity among subjects with a non-pCR was 376 times higher than the odds for positivity among subjects with a pCR. In the same way, ∆skewness and post-nCRT skewness demonstrated respectable diagnostic performances with DOR values of 192.17 and 168.56, respectively. Although these values appear to be lower than DORs of ∆%kurtosis and post-nCRT kurtosis, the confidence intervals for these metrics clearly overlap, so we cannot conclude that the kurtosis-related parameters were statistically better than the skewness-related parameters using DOR.

Finally, this study confirm that ADC histogram analysis is a reproducible technique. Similarly, van Heeswijk et al[49] demonstrated that histogram-derived parameters had good interobserver agreement, with ICC values ranging from 0.80-0.98. This result supports the method’s validity and suggests that it can be used in clinical practice. Furthermore, we utilized non-precise tumor delineation, which was quicker and produced comparable findings to those obtained by an expert radiologist's measurement, suggesting that this technique could be performed semiautomatically with an excellent interobserver agreement. This finding is very important when considering the implementation of histogram analysis in routine clinical practice.

Our study had important limitations. First, this was a retrospective, single-center evaluation. We believe that the present study might serve as a foundation for larger prospective studies in the future. Second, we included only a small number of patients (n = 48), while no validation group was included (both restricting the conception of a predictive model by using a multivariate logistic regression analysis). Third, the patient numbers among the different histopathologic TRGs were not well balanced. Only 18 patients (38%) achieved a histopathologic complete response, which may have introduced an element of statistical bias. However, these patients achieved a strict pCR, underlying the high degree of accuracy of our metrics. Fourth, the parameters obtained from the hotspot ROI were not conclusive enough to predict treatment response in the present study. This result is still in significant disagreement with our prior work where we demonstrated a high diagnostic accuracy of the ∆%ADCmean when distinguishing a pCR in rectal cancer by choosing a cutoff value of 55%[26]. Differences in research methods might explain this discrepancy, but we sustain that it is more reliable to use volumetric ROIs than one slice ROIs.

In summary, although further studies are needed to address the limitations of the current work, we demonstrated the benefits of considering measures other than the ROI ADCmean to evaluate the response to therapy in patients with LARC. Moreover, kurtosis and skewness have been selected by many radiomics studies of rectal cancer, emphasizing the importance of first-order statistics features for the assessment of therapy response[47,50]. Our results support the importance of these parameters, but they also helped us to standardize both the extraction and analysis of the data collected, which is a crucial step when developing and validating our own multiparametric model to predict treatment outcomes.

CONCLUSION

Based on the DWI technique, some whole-lesion histogram parameters could provide valuable information when diagnosing rectal cancer. In particular, kurtosis and skewness might be a useful indicator in the preoperative evaluation of a pCR in rectal cancer. Understanding skewness and kurtosis of the ADC parameters is the simplest way to recognize the deviation of Gaussianity, which indicates tumor heterogeneity. Moreover, we demonstrated high interobserver reliability for measurements of all of the histogram-derived parameters analyzed in the current work, addressing the challenges associated with replication that are well-known among more complex predictive models. Further long-term studies are needed to determine the ultimate clinical utility of our results.

ARTICLE HIGHLIGHTS

Research background

Studies have shown that successful treatment of many tumors can be detected using diffusion-weighted magnetic resonance imaging (MRI) as an increase in the apparent diffusion coefficient (ADC). However, findings from rectal cancer have been limited. Therefore, the criteria used for tumor staging and surveillance are largely based on anatomic criteria at this time. Broadly, whole lesion histogram analysis of ADC aims to fill this gap, extracting and analyzing the higher quantitative data with the aim of more accurate, tumor-specific evaluation and characterization.

Research motivation

ADC histogram parameters reflect the distribution and variation of all voxels within the entire lesion, which reduce the subjectivity of region of interest (ROI) placement and improves repeatability in the quantitative ADC analysis. Previous studies have applied volumetric ADC histogram analysis to predict treatment response of squamous carcinoma, breast cancers, and ovarian cancers. No ADC histogram study thus far has focused on locally advanced rectal cancer (LARC).

Research objectives

We aim to evaluate the effectiveness of whole lesion histogram analysis of ADC in the prediction to neoadjuvant chemoradiation therapy (nCRT) response in patients with LARC.

Research methods

This was a retrospective study. We collected data of 48 consecutive patients with histologically confirmed LARC. All patients underwent a pre-treatment MRI for primary tumor staging and a second restaging MRI for response evaluation. The sample was distributed as follows: responders (R), n = 18; and non-responders (non-R), n = 30. Eight parameters derived from the histogram analysis of ADC, as well as the ADCmean from the hot spot ROI, were obtained and compared between R and non-R. The diagnostic accuracy in the prediction of treatment response of all variables included in the present study was calculated as well.

Research results

Post-nCRT kurtosis, ∆%kurtosis, post-nCRT skewness an ∆skewness exhibited the highest diagnostic performance in predicting a good response to nCRT.

Research conclusions

The results of our study support that histogram-parameters derived from ADC values can be used to stratify good responders into studies exploring individualized, less extensive treatment regimens, such as the omission of radiotherapy and less extensive surgery, or even deferral of surgery.

Research perspectives

We need to expand the sample size to confirm further the diagnostic accuracy of kurtosis and skewness. In addition, the long-term outcome of this analysis should be a radiomic model for predict treatment response in rectal cancer.