SA-Net: A scale-attention network for medical image segmentation.

Semantic segmentation of medical images provides an important cornerstone for subsequent tasks of image analysis and understanding. With rapid advancements in deep learning methods, conventional U-Net segmentation networks have been applied in many fields. Based on exploratory experiments, features at multiple scales have been found to be of great importance for the segmentation of medical images. In this paper, we propose a scale-attention deep learning network (SA-Net), which extracts features of different scales in a residual module and uses an attention module to enforce the scale-attention capability. SA-Net can better learn the multi-scale features and achieve more accurate segmentation for different medical image. In addition, this work validates the proposed method across multiple datasets. The experiment results show SA-Net achieves excellent performances in the applications of vessel detection in retinal images, lung segmentation, artery/vein(A/V) classification in retinal images and blastocyst segmentation. To facilitate SA-Net utilization by the scientific community, the code implementation will be made publicly available.

Introduction

Since manual and dense labeling of a large number of medical images is time-consuming, tedious and prone to inter- and intra-observers, automatic methods for medical image segmentation have been rapidly emerging. These methods lead to accurate and reliable solutions that could improve clinical workflow efficiency and support health-care decision making by allowing quick and automatic extraction of useful quantitative information.

Good representation capabilities, efficient inference, and filter sharing make convolutional neural networks (CNN) the de facto standard for image segmentation. Full convolutional networks (FCN) [1] demonstrated good semantic segmentation performance on the Pascal VOC dataset. U-Net [2] and U-Net variants models have been successfully used in segmenting biomedical images of neuronal structures. In particular, MultiResUNet [3] achieved the best segmentation accuracy for neuronal structures in electron microscopy. Recently, both FCN and U-Net have achieved semantic segmentation performance that closely matches the radiologist performance in many tasks [4-10].

In recent years, many U-Net variants have been devised for different tasks of medical image segmentation. Fu et al. [4] adopted a conditional random field (CRF) to extract multi-stage features for improving vessel detection outcomes. The M-Net [5] architecture, a variant of U-Net, was proposed for joint segmentation of the optic disc and cup by augmenting U-Net with deep supervision and multi-scale inputs. Alom et al. [6] proposed RU-Net, a U-Net variant equipped with recurrent convolution. Ozan et al. [7] improved the U-Net performance with an attention mechanism. Simon et al. [8] proposed the Tiramisu architecture in which dense blocks convolutions replace the U-Net convolutional layers. Other CNN variants, such as PSPNet [9] and DeepLab [10], were introduced to achieve superior performance on benchmark tasks of semantic segmentation. Despite the emergence of the afore-mentioned variants, U-Net is still the most common architecture for medical image segmentation, essentially as its encoder-decoder organization (together with its skip connections) does not hinder efficient information flow, and its performance does not deteriorate at low data regime.

U-Net and U-Net like models has been showing impressive potential in segmenting medical images, but the performance of these models will be poor when the target organ exhibits large shape and size variations among patients. Therefore, design good multi-scale features for medical images segmentation is essential. However, creating multi-scale representations requires feature extractors to use receptive fields of considerable variations to give a detailed account of parts, objects, or context at all possible scales. The natural way for CNNs to extract coarse-to-fine multi-scale features is to utilize a convolutional operator stack. Such inherent CNN capability of extracting multi-scale features leads to good representations for handling numerous medical image analysis tasks.

To handle the problem of scale variations, Adelson et al. [11] intuitively leveraged multi-scale image pyramids, and a technique that is quite common in approaches based on hand-crafted features [12, 13] and CNN features. There is a concrete evidence [14, 15] that multi-scale feature learning could be beneficial for deep-learning detectors [16, 17]. The sensitive nonlinear iterative peak (SNIP) algorithm [18, 19] achieves scale normalization by selectively picking training objects of suitable dimensions for each image scale. This algorithm avoids objects of extreme scales, i.e., small or large objects under relatively smaller or larger scales, respectively. However, the computationally high inference times of the image pyramid methods make these methods practically infeasible. The CE-Net [20] architecture employs Dense Atrous Convolution (DAC) blocks to create a multi-scale network for better medical image understanding. Atrous/Dilated convolution [10] expands convolutional kernels by carrying on convolution at sparsely sampled positions. Dilated convolution is frequently utilized in semantic segmentation to account for large-scale contextual information [21, 22]. However, it still suffers from some potential shortcomings, for example, it may cause some pixels never participate in the calculation, which is not friendly to pixel-level prediction. In addition, although the dilated convolution guarantees a larger receptive field with no additional parameters, it is extremely unfriendly for some small objects that do not need such a large receptive field. Moreover, in comparison to the conventional FCN, ResNet-101 [23] has 23 residual blocks (with 69 convolutional layers) of Dilated FCN which require 4 times more computational operations and memory resources, while 3 residual blocks (with 9 convolutional layers) need 16 times more resources. Recently, Res2Net [24] has been constructed as individual residual blocks where each block has hierarchical residual connections. Res2Net adopts a granular-level representation of multi-scale features and enlarges the receptive field range for every network layer. Yet, without exploiting information at different scales, many pieces of redundant information are also transmitted to large-scale features.

Motivated by the afore-mentioned approaches, we make the following key contributions:

We propose a new scale-attention deep learning network (SA-Net) based on the residual module and attention module, which could segment the different medical images effectively.

To capture more multi-scale features to better comprehend the structure and function of different tissues in a medical image, an effective Scale-Attention (SA) module is introduced.

The proposed method is tested in lung segmentation, retinal vessel detection, artery/vein (A/V) classification and blastocyst segmentation tasks. The experimental results demonstrate a superior performance on various tasks compared to the competing methods.

The remainder of the paper is structured as follows. We outline the proposed deep learning framework in Section Methods. The experimental settings are described, and the results and discussion are presented in Section Experimental evaluation and results. Our key conclusions are stated in Section Discussion and conclusions.

Methods

This section presents in detail the design of the scale-attention network for medical image segmentation. Firstly, we use a basic U-Net as the backbone network. In addition, we insert a SA module in the bridge connection of the U-Net, which could extract multi-scale residual features for achieve our aim of a scale-attention network for different tissues segmentation in medical images. Fig 1 outlines the SA-Net framework.

Fig 1

Diagram of the proposed SA-Net.

Our main target herein is to obtain a feature map that can be learned to integrate different scale representations according to different tissue scales in the input medical images. For example, in the retinal fundus images (Fig 2), the main blood vessels branch into micro blood vessels, and the blood vessel diameters in the image range from 1 to 35 pixel units. Fig 2 shows that the micro blood vessels are of very high frequency. Therefore, understanding the huge scale variations are quite crucial and challenging.

Fig 2

Visualization of the retinal vessel diameters in the fundus.

(a) The raw 1880×2886 image. (b) Diameter of each point on the skeleton. c) Partially enlarged view of the red area in (b). (d) Histogram of the diameter map for each point on the skeleton (distance between pixels).

Multi-scale features

In this work, we use the capabilities of Res2Net [24] to learn and understand the image features at different scales. Instead of extracting features using a group of 3×3 filters as in the ResNet [23] bottleneck block (as shown in Fig 3(A)), we propose a Res2Net variant with better extraction capability of multi-scale features, with roughly the same computational cost. The 3×3 filter groups are replaced with smaller filter groups connected in a hierarchical residual-type manner. As shown in Fig 3(B), after the 1×1 convolution, the features are split into k subsets, denoted by x, where i∈{1,2,…,k}. While all subsets have the same spatial size, the channel count for each subset is 1/k times that of the input feature map. Each subset x (except for x1) has a 3×3 convolution filter F().

Fig 3

Comparing the Res2Net and ResNet blocks (with a scale dimension of k = 4): (a) The conventional building block in CNN variants. (b) Res2Net uses a group of 3×3 filters. (c) The SA module adds the attention module to enforce the scale-attention capability.

In SA module, we add the attention module to enforce the scale-attention capability. The details of proposed attention module are shown in Fig 4. Firstly, a max-pooling and an average-pooling are used to obtain the global information of each channel, which automatically highlight the relevant feature channels while suppressing irrelevant channels. Then, their results are summed and fed into a 1×1 convolutional layer followed by the sigmoid activation function. Finally, the output result is obtained by multiplying with the input.

Fig 4

Diagram of the attention module.

As illustrated, the attention module utilizes both max-pooling outputs and average-pooling outputs.

Scale-attention module

For the best possible transfer of the useful small-scale field-of-view features to large-scale features, we propose a scale-aware (SA) module, as shown in Fig 3(C). This SA module adds the attention model (A()) with the argument y, as shown in Fig 4. First, we get the attention map A(y) and concatenate it with x, and then feed the result into F(). To decrease the parameters while allowing the number of subsets k to increase, we omit the 3×3 convolution for x1 herein. Thus, y can be written as:

Each 3×3 convolutional operator F() might get information from all feature subsets {x, j≤i}. When a feature subset x is processed by a 3×3 convolutional operator, the resulting output may have an enlarged receptive field compared to x. The high combinatorial complexity causes the SA module output to have different numbers and combinations of the receptive field size and scales.

We process the feature subsets in the SA module following a multi-scale approach, where local and global information is extracted. All subsets are concatenated and processed with 1×1 convolution, with the aim of achieving better fusion of information at different scales. More effective feature convolution is achieved by this split and concatenation strategy. With the objective of reducing the parameter count, a feature reuse approach is followed where the convolution for the first subset was omitted.

The scale dimension k is used herein as a control parameter. A larger k value might enable the learning of features with richer receptive field sizes, with insignificant concatenation-induced overheads in terms of computations and memory usage.

The loss function

We employ an end-to-end deep-learning scheme as our underlying framework. Fig 1 asserts the need for training the proposed system to predict the segmentation label of each pixel. The loss is quantified by the commonly-used cross-entropy loss function. while the total loss is defined as: where n denotes the number of pixels in the input image, y′ is the predicted output probability of a foreground pixel, and y is the ground-truth pixel. We use L2 regularization with a weight of β = 0.0002.

Experimental evaluation and results

Experimental settings

This section introduces the image preprocessing and data augmentation procedures for network training. And also provides ample details about our experimental setup.

Data transformations and augmentation are needed during training to avoid model overfitting. In medical imaging, an essential constraint of these transformations is that the output images must be quite realistic. To increase the training data variability while achieving this realism, we employ only two-dimensional rotations (through random angles) of each training batch. In particularly, we realize that the background color and illumination of different fundus images induce a large variability in pixel intensity. This variability is inherent in the training data. Contrast enhancement could be used to reduce this variability and increase the image quality in image preprocessing. And the impact of individual differences is avoided by using gray retinal images instead of color images [2, 25]. In addition, in this study we select the last epoch when the loss of the training model fluctuates less than 0.01 within 20 epochs as our final model for testing.

Our system was implemented using an Ubuntu 16.04 operating system, an Intel® Xeon® Gold 6148 CPU with a 2.40-GHz processor and 256-GB RAM, NVIDIA Tesla V100 GPU, a PyTorch backend, and cuDNN 9.0.

Retinal vessel detection

For retinal vessel detection in fundus images, we evaluated our method on two publicly available datasets. Firstly, we used the 40-images DRIVE [26] dataset, for which two expert human annotations are available. The first annotation is typically used as the gold standard [4]. The DRIVE dataset consists of 20 training images and 20 testing images with the resolution of 584×565. Secondly, we examined the 28-images CHASE_DB1 database [27], with images from both eye sides for each of 14 children. A clear discrimination couldn’t be reached between the healthy and diseased cases in CHASE_DB1. Thus, we adopted a scheme of stratified k-fold cross-validation, in which the input data is subdivided into equally-sized k folds, such that one fold is set for testing, while the other (k-1) folds are set for training. The results of k repetitions of this process are averaged to get pooled estimates of segmentation metrics. For performance consistency among cross-validation folds, the same settings and training initialization were used. For our CHASE_DB1 experiment, four folds were used with 7 images each divided almost evenly among the two eye sides. The ground-truth segmentation data for CHASE_DB1 was generated using Hoover’s annotations. For the DRIVE images, binary segmentation masks are publicly available. Field-of-view (FOV) masks for CHASEE_DB1 were manually created following [28]. Example images and masks for this two datasets are demonstrated in Fig 5.

Fig 5

Samples of the retinal fundus images and their H×W dimensions in pixels.

(a) Input images: Left: DRIVE (584×565); right: CHASE_DB1 (960×999). (b) Manual annotation of the retinal vessels. (c) Binary segmentation masks of the fundus images.

For blood vessel detection performance evaluation, several metrics were computed, namely sensitivity (SE), specificity (SP), accuracy (ACC), Matthews correlation coefficient (MCC), and the F1 score (F1) [4]. Moreover, for the receiver operating characteristic curve, we compute the area under the curve (AUC) and use it as well to assess the segmentation performance. For all of these metrics, a perfect detector gives a value of 1. A threshold value of 0.5 was applied to the probability maps to obtain the binary segmentation outputs. Only pixels inside the field of view were processed.

The performance of SA-Net was compared against relevant algorithms [29-31], in addition to recently-developed deep learning methods [2, 6, 25, 32–35]. Table 1 shows the comparative experimental results. Indeed, for DRIVE dataset, SA-Net achieves remarkable performance in terms of the F1, MCC, SE and AUC metrics, with values of 0.8289, 0.8055, 0.8252 and 0.9822, respectively. For CHASE_DB1, all of the MCC, SE, ACC, AUC and F1 metrics achieve excellent performance, with values of 0.8102, 0.8199, 0.9665, 0.9865 and 0.8280, respectively. For visualization results, example outputs are given in Fig 9. For the identification of the blood vessels at different scales, the SA-Net shows more continuous results and the detection of small blood vessels are especially close to the true labels. These results clearly indicate that devising the SA-Net architecture with multi-level scale-aware capabilities led to superior retinal vessel segmentation performance.

Table 1

Segmentation performance metrics on two publically available retinal image datasets.

Datasets		Method	MCC	SE	SP	ACC	AUC	F1
DRIVE	Unsupervised	Zhao [29]	N/A	0.7420	0.9820	0.9540	0.8620	N/A
		Azzopardi [30]	N/A	0.7655	0.9704	0.9442	0.9614	N/A
		Roychowdhury [31]	N/A	0.7395	0.9782	0.9494	0.8672	N/A
	Supervised	U-Net [2]	N/A	0.7537	0.9820	0.9531	0.9755	0.8142
		RU-Net [6]	N/A	0.7792	0.9813	0.9556	0.9784	0.8171
		DE-Unet [25]	N/A	0.7940	0.9816	0.9567	0.9772	0.8270
		SWT-UNet [32]	0.8045	0.8039	0.9804	0.9576	0.9821	0.8281
		BTS-UNet [33]	0.7923	0.7800	0.9806	0.9551	0.9796	0.8208
		Driu [34]	0.7941	0.7855	0.9799	0.9552	0.9793	0.8220
		CS-Net [35]	N/A	0.8170	0.9854	0.9632	0.9798	N/A
		SA-Net (Ours)	0.8055	0.8252	0.9764	0.9569	0.9822	0.8289
CHASE-DB1	Unsupervised	Azzopardi [30]	N/A	0.7585	0.9587	0.9387	0.9487	N/A
		Roychowdhury [31]	N/A	0.7615	0.9575	0.9467	0.9623	N/A
	Supervised	RU-Net [6]	N/A	0.7756	0.9820	0.9634	0.9815	0.7928
		SWT-UNet [32]	0.8011	0.7779	0.9864	0.9653	0.9855	0.8188
		BTS-UNet [33]	0.7733	0.7888	0.9801	0.9627	0.9840	0.7983
		SA-Net (Ours)	0.8102	0.8199	0.9827	0.9665	0.9865	0.8280

aN/A = Not Available, SWT-UNet shows the vessel segmentation results using fully convolutional neural networks, BTS-DSN gives the segmentation results with the multi-scale deeply-supervised networks with short connections.

Fig 9

Example results for lung segmentation, detection of retinal blood vessels, artery/vein classification and blastocyst segmentation.

From top to bottom: lung segmentation, retinal vessel detection, artery/vein classification and blastocyst segmentation.

Lung segmentation

In this section, we seek to segment the lung structures in 2D CT images. We evaluate SA-Net on 2D lung CT images provided by the Lung Nodule Analysis (LUNA) competition, in which two challenges have been made, namely nodule detection and false-positive reduction. The LUNA dataset contains 534 publically-available images of 512×512 pixels each, and corresponding segmentation masks [36]. We experiment with a cross-validation scheme in addition to a scheme with 80% training images and 20% testing ones, which same with the CE-Net [20]. Sample CT images are displayed in Fig 6.

Fig 6

Sample lung CT images with dimensions of H×W in pixels.

The performance metrics used herein are the overlapping error E (defined below), the accuracy and the sensitivity [20]. In addition to the average metric values are reported in Table 2. The overlapping error is given by where P and T denote the predicted and ground-truth lung segmentations, respectively.

Table 2

Segmentation performance measures for lung image datasets.

Method	E	ACC	SE
U-Net [2]	0.087	0.975	0.938
CE-Net [20]	0.038	0.990	0.980
SA-Net (Ours)	0.035	0.986	0.988

From Table 2, as we can see that SA-Net reaches an overlapping error of 0.035, a sensitivity of 0.988, and an accuracy of 0.986. These values are better than the corresponding ones obtained by U-Net. Also, we compare SA-Net with the CE-Net [20] architecture, which pays more attention to high features. The overlapping error drops by 7.89% from 0.038 to 0.035, while the sensitivity increases from 0.980 to 0.988. Some examples are also given in Fig 9. As shown, because of the addition of the SA module, the lung structure can be better identified, and the segmentation result is more consistent than that of U-Net (which wrongly identifies non-lung tissues as lung ones). The results further emphasize the significance of our proposed SA blocks for lung segmentation.

Artery/Vein classification

Our third application is A/V classification in fundus images. Motivated by the lack of a gold standard that allows objective comparison of approaches for this problem, Qureshi et al. [37] created a manually-annotated benchmark for A/V classification using the DRIVE dataset. The labeling was performed by one ophthalmologist and two computer vision scientists. A majority vote among the three experts was used to decide the ground-truth blood vessel class. Also, as described by Hu et al. [38], ground-truth data was produced from the binary vessel segmentation created by the second expert. In this work, the ground-truth acts as a proxy for A/V classification by an independent expert. The training and testing subsets were obtained following the same schemes as those of the blood vessel detection problem, which means that 20 images are used for training and the remaining 20 are used for testing. The Balance accuracy (BACC), SEAV, SPAV, F1 score of arteries (F1A) and F1 score of veins (F1V) are used in this section to evaluate the SA-Net performance. The image and label samples are displayed in Fig 7.

Fig 7

A sample retinal image from the DRIVE dataset and its corresponding artery/vein segmentation map.

Where TP is the number of artery pixels correctly classified, TN is the number of vein pixels correctly classified, FP is the number of vein pixels mis-classified as artery pixels and FN is the number of artery pixels mis-classified as vein pixels.

In Table 3, the SA-Net result shows better performance. This demonstrates the validity of our model. As well, it can be seen that the results for veins are better than those for arteries. This difference is mainly because of the clinical manifestations, i.e., the color of the vein is dark red, while the arterial blood is light red. Some examples are given in Fig 9. We can see that the identification of blood vessels on different scales is better whether these vessels are veins or arteries. Because of the addition of the SA module, SA-Net works better in distinguishing arteries and veins, while the confusion between arteries and veins is serious for U-Net. These results reaffirm the superiority of the SA module in learning images with different scales.

Table 3

Performances of different A/V classification methods on DRIVE dataset.

Method	BACC	SEAV	SPAV	F1A	F1V
U-Net [2]	0.9122	0.9145	0.9083	0.7089	0.7586
DoS [39]	N/A	0.9190	0.9150	N/A	N/A
SA-Net (Ours)	0.9351	0.9345	0.9347	0.7336	0.7802

Blastocyst segmentation

The fourth application is blastocyst images segmentation. There is a detail introduction to the human embryo dataset in [40], which has been open sourced. This blastocyst dataset contains 235 blastocyst images collected from time-lapse. Various tissue labels are provided by the Pacific Centre for Reproductive Medicine (PCRM). The training set accounts for 85% of all data while the test set accounts for the remaining 15%. And the division of the training set and test set is consistent with [41]. The sample is shown in Fig 8.

Fig 8

A sample blastocyst image from the blastocyst dataset and its manual label.

For blastocyst images segmentation performance evaluation, Jaccard Index, also called Intersection-over-Union, is utilized to evaluate segmentation performance. The metrics is same with the Blast-Net [41].

Our method is compared with several classic methods [9, 10, 41, 42]. Table 4 shows the comparative experimental results. As for Inner Cell Mass (ICM), Blastocoel, Trophectoderm (TE), Zona Pellucida (ZP) and Background segmentation, it can be seen that SA-Net have achieved significant improvement performance, which are 83.67%, 89.56%, 77.50%, 90.93% and 97.52%, respectively. Moreover, from Fig 9 we can see that SA-Net shows better segmentation performance compared with the basic U-Net. These results also have proved that the addition of the SA model has better comprehension for the multi-scale structural features of different organs.

Table 4

Blastocyst segmentation performance measures on the blastocyst dataset.

Method	Mean	Background	Blastocoel	ZP	TE	ICM
U-Net [2]	0.8137	0.9404	0.7941	0.7932	0.7506	0.7903
PSPNet [9]	0.8151	0.9460	0.7926	0.8057	0.7483	0.7828
DeepLab V3 [21]	0.8165	0.9449	0.7835	0.8084	0.7398	0.8060
Blast-Net [41]	0.8285	0.9474	0.8079	0.8115	0.7652	0.8107
TernausNet [42]	0.8142	0.9450	0.7861	0.8024	0.7616	0.7758
SA-Net (Ours)	0.8783	0.9752	0.8956	0.9093	0.7750	0.8367

Ablation study

For further assessment of the SA module, we carried on an ablation study on the DRIVE dataset. The U-Net architecture was used as the baseline model. In one variant, we replaced the middle skip connections with the bottleneck and ResNet, and with the bottleneck and the SA module for Res2Net. We then verified the importance of the multiscale and attentional features.

In Table 5, adding the residual module increases the ACC, AUC and F1 scores to 0.9566, 0.9817 and 0.8244, respectively. We can see that the residual module is indeed very helpful for improving the learning characteristics. When adding the multi-scale features to the Res2Net learning, it can be found that the sensitivity and F1 scores had a significant increase, while the other indicators showed no obvious improvements, and even the AUC metric slightly decreased. After the introduction of the multi-scale scheme, the model ability to understand the scale features became stronger and the sensitivity became higher. However, some useless information was inadvertently introduced, resulting in feature redundancy. When we added our proposed SA module, we can see that all indicators except for the specificity have increased, among which MCC, SE, ACC, AUC and F1 scores increase to 0.8055, 0.8252, 0.9569, 0.9822, and 0.8289, respectively. These results demonstrate the effectiveness of the proposed scale-attention module to a certain extent.

Table 5

Ablation study segmentation results on the DRIVE dataset.

Method	MCC	SE	SP	ACC	AUC	F1	Params
Backbone [2]	N/A	0.7537	0.9820	0.9531	0.9755	0.8142	34.5M
Backbone+ResNet	0.8013	0.8039	0.9793	0.9566	0.9817	0.8244	49.5M
Backbone+Res2Net	0.8026	0.8148	0.9775	0.9566	0.9816	0.8260	161M
Backbone+SA	0.8055	0.8252	0.9764	0.9569	0.9822	0.8289	194M

Discussion and conclusions

The task of segmenting different tissue types in medical imaging is very critical, and good segmentation results are conducive to the analysis of some tasks in later stages. In this paper, we proposed an end-to-end scale-attention deep learning network, and experiment results show that SA-Net has achieved superior performance results in retinal vessel detection, lung segmentation, artery/vein classification and blastocyst segmentation tasks. More importantly, it can be seen from the visualization results in Fig 9 that SA-Net shows better performance for segmentation of different medical images with different scales tissues compared with traditional U-Net architecture, indicating that SA-Net can better learn the features at different scales. In this work, the proposed SA-Net achieves effective segmentation in various medical images, it can be considered as a better model for 2D small-sample medical image segmentation. While we only explored 2D medical images in this work, we will explore the application of SA-Net for segmenting 3D medical images and consider reducing the computational cost of the model in the future.

17 Nov 2020

PONE-D-20-30729

SA-Net: A Scale-Attention Network for Medical Image Segmentation

PLOS ONE

Dear Dr. Zhang,

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 submit your revised manuscript by Jan 01 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

We look forward to receiving your revised manuscript.

Kind regards,

Yuankai Huo, Ph.D.

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 suggest you thoroughly copyedit your manuscript for language usage, spelling, and grammar. If you do not know anyone who can help you do this, you may wish to consider employing a professional scientific editing service.

Whilst you may use any professional scientific editing service of your choice, PLOS has partnered with both American Journal Experts (AJE) and Editage to provide discounted services to PLOS authors. Both organizations have experience helping authors meet PLOS guidelines and can provide language editing, translation, manuscript formatting, and figure formatting to ensure your manuscript meets our submission guidelines. To take advantage of our partnership with AJE, visit the AJE website (http://learn.aje.com/plos/) for a 15% discount off AJE services. To take advantage of our partnership with Editage, visit the Editage website (www.editage.com) and enter referral code PLOSEDIT for a 15% discount off Editage services.  If the PLOS editorial team finds any language issues in text that either AJE or Editage has edited, the service provider will re-edit the text for free.

Upon resubmission, please provide the following:

The name of the colleague or the details of the professional service that edited your manuscript

A copy of your manuscript showing your changes by either highlighting them or using track changes (uploaded as a *supporting information* file)

A clean copy of the edited manuscript (uploaded as the new *manuscript* file)

3. Thank you for stating the following in the Acknowledgments Section of your manuscript:

"This research was supported by Hefei Innovation Research Institute, Beihang University and ‘the Thousand Talents Plan’ Workstation between Beihang University and Jiangsu Yuwell Medical Equipment and Supply Co. Ltd. The authors are grateful to all study participants."

We note that you have provided funding information that is not currently declared in your Funding Statement. However, funding information should not appear in the Acknowledgments section or other areas of your manuscript. We will only publish funding information present in the Funding Statement section of the online submission form.

Please remove any funding-related text from the manuscript and let us know how you would like to update your Funding Statement. Currently, your Funding Statement reads as follows:

"This work was supported by the National Key Research and Development Program of China under Grant 2016YFF0201002, the National Natural Science Foundation of China under Grant 61572055, the University Synergy Innovation Program of Anhui Province GXXT-2019-044. These awards were received by Jicong Zhang. YES - Specify the role(s) played."

Additionally, because some of your funding information pertains to commercial funding, we ask you to provide an updated Competing Interests statement, declaring all sources of commercial funding.

In your Competing Interests statement, please confirm that your commercial funding does not alter your adherence to PLOS ONE Editorial policies and criteria by including the following statement: "This does not alter our adherence to PLOS ONE policies on sharing data and materials.” as detailed online in our guide for authors  http://journals.plos.org/plosone/s/competing-interests.  If this statement is not true and your adherence to PLOS policies on sharing data and materials is altered, please explain how.

Please include the updated Competing Interests Statement and Funding Statement in your cover letter. We will change the online submission form on your behalf.

[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: Yes

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: This paper proposed a scale-attention network upon the backbone network of UNet. And designed multiple attention modulus at different scales, on segmentation tasks. The attention mechanism at scales improved the extraction of global-local features. The proposed method is trained and tested on multiple datasets including two public retina vessel datasets and the LUNA. The method is also validated on the artery/vein classification problem and blastocyst segmentation. The study across multiple datasets are impressive. In addition, the method is compared with extensive baseline methods in Table 1, which is a bonus and provides a good reference for future studies.

Overall, this is a well evaluated paper with multiple applications and large-scale experiments. However, I have several concerns about several places, especially in the description of abstract/introduction and discussion. It would be great to further tune some statements, please see as follows:

1. The abstract address that “uses an attention module to learn and understand which features are the most important for medical image segmentation”. However, this paper didn’t evaluate the feature importance or have the conclusion on which features are the most important. I would recommend the authors to add the evaluation of different features effects or to rephrase the sentence.

2. I suggest changing the description of “not only a waste of manpower and time but also prone” in the first sentence of introduction to “manual effort is time-consuming and tedious”, we should be humble to previous manual work by radiologists and clinicians, these are hard tasks and defines many critical medical problems.

3. “In particular, U-Net [2] achieved the best segmentation accuracy for neuronal structures in electron microscopy.” This claim lacks a citation, could the authors add the citation on the best performed paper of “neuronal structures in electron microscopy”?\\

4. Suggestion to reduce redundancy: “and its performance does not deteriorate even when large enough datasets are lacking” to “and its performance does not deteriorate at low data regime”, many other places are similar, it will be great if the authors could further read and try reduce redundancy.

5. The fourth paragraph in Introduction, “Nevertheless, …., these variants still rely on cascaded multi-stage CNNs. Therefore, we emphasize the design of particular good multi-scale features”. This sentence is confusing, I believe the propose method is still based on encoder-decoder architecture and used the cascaded multi-stage CNNs. Could the authors rephrase the sentence and highlight the difference between the proposed work and convention UNet?

6. The legend of Figure 1 and Figure 2 are too small to see, could enlarge the legend.

7. The experiment setting separated the datasets into training and testing or cross-validation, since there is no validation set, how the final model is selected for testing?

8. Followed by the first concern, the authors highlighted the ability to “better learn and understand the features at different scales” at the Discussion and Conclusion, it would be more intuitive to visualize the attention maps at different scales for reader, and discuss the difference with non-multi-scale attention. Otherwise, I would suggest removing this statement.

Summary:

This is a well-evaluated paper. The large-scale experiments and comparisons are impressive. The application on five different datasets is a bonus. Overall, this is a great study, it should be of potential interest of readers if authors could address above issues and further fine-tune some descriptions.

Reviewer #2: Overall, this work proposes a scale-attention module to the Res2Net based UNet for the task of classification and segmentation in the medical imaging field. Experiments are done on 4 publicly available datasets and with one ablation study.

Cons & Questions:

1) The claimed contribution is the proposed scale-attention module. However, according to the ablation study in Table V, the improvement of SA compared with Res2Net is really marginal, only 0.03% in ACC, 0.06% in AUC and 0.29% in F1, which could also be in the std range if running different algorithms for multiple times. Also, the analysis in the main text "In fact, the MCC, SE, ACC, AUC and F1 scores increase by 0.36%, 1.28%, 0.03%, 0.06%, and 0.35%, respectively" is not consistent with the numbers in Table V. Can authors check that?

2) What is the computation cost comparison for the ablation study in Table V? I can image a marginal improvement can be obtained by a bit more computation cost.

3) Some important details of some datasets are missing. The training and testing split of the DRIVE dataset and the A/V classification dataset are missing. What is the GPU version you are using for the training and testing?

4) In table IV, the reference number of Blast-Net is mixed to be the TernausNet paper.

Pros:

1) The written is clear. The scale problem is a pain point in the medical imaging, even for the computer vision.

2) The experiments are done on many datasets.

**********

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: No

Reviewer #2: No

[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.

Submitted filename: Reviewer_plosOne_SANet.pdf

Click here for additional data file.

24 Dec 2020

Dear Editors and Reviewers:

We thank you very much for giving us an opportunity to revise our manuscript. Thanks a lot for the helpful comments and recommends and thank you for your time spent. According to the reviewers' comments, we have revised the manuscript. Revised portion are marked in red in the paper.

The main corrections in the paper and the responds to the comments are as follows:

Reviewer #1:

General comments

This paper proposed a scale-attention network upon the backbone network of UNet. And designed multiple attention modulus at different scales, on segmentation tasks. The attention mechanism at scales improved the extraction of global-local features. The proposed method is trained and tested on multiple datasets including two public retina vessel datasets and the LUNA. The method is also validated on the artery/vein classification problem and blastocyst segmentation. The study across multiple datasets are impressive. In addition, the method is compared with extensive baseline methods in Table 1, which is a bonus and provides a good reference for future studies.

Overall, this is a well evaluated paper with multiple applications and large-scale experiments. However, I have several concerns about several places, especially in the description of abstract/introduction and discussion. It would be great to further tune some statements, please see as follows:

1. The abstract address that “uses an attention module to learn and understand which features are the most important for medical image segmentation”. However, this paper didn’t evaluate the feature importance or have the conclusion on which features are the most important. I would recommend the authors to add the evaluation of different features effects or to rephrase the sentence.
Response: Thanks for your suggestion. Indeed, our description is confusing. To avoid confusion, we have rephrased the sentence “uses an attention module to learn and understand which features are the most important for medical image segmentation” in the abstract as “uses an attention module to enforce the scale-attention capability. SA-Net can better learn the multi-scale features and achieve more accurate segmentation for different medical image.” according to the reviewer’s comment.

2. I suggest changing the description of “not only a waste of manpower and time but also prone” in the first sentence of introduction to “manual effort is time-consuming and tedious”, we should be humble to previous manual work by radiologists and clinicians, these are hard tasks and defines many critical medical problems.
Response: Thanks for the suggestion. We think the reviewer’s comment is very reasonable and we are also very ashamed of the inappropriate expression. Our description is indeed inappropriate, we have revised the description of “not only a waste of manpower and time but also prone” in the first sentence of introduction according to the reviewer’s suggestion. As shown in the first sentence of Introduction marked in red.

3. “In particular, U-Net [2] achieved the best segmentation accuracy for neuronal structures in electron microscopy.” This claim lacks a citation, could the authors add the citation on the best performed paper of “neuronal structures in electron microscopy”?
Response: Thanks for pointing it out. We are very sorry that we ignored this citation, in the revised version, we have added reference [3]：“Ibtehaz N, Rahman M S. MultiResUNet: Rethinking the U-Net architecture for multimodal biomedical image segmentation[J]. Neural Networks, 2020, 121: 74-87.”, which is a reference on the best performed paper of “neuronal structures in electron microscopy”. The revised version is “U-Net [2] and U-Net variants models have been successfully used in segmenting biomedical images of neuronal structures. In particular, MultiResUNet [3] achieved the best segmentation accuracy for neuronal structures in electron microscopy.”

4. Suggestion to reduce redundancy: “and its performance does not deteriorate even when large enough datasets are lacking” to “and its performance does not deteriorate at low data regime”, many other places are similar, it will be great if the authors could further read and try reduce redundancy.

Response: Thanks for the suggestion. In order to reduce redundancy, we have carefully read and revised the whole manuscript according to the reviewer’s comment. The revised place has been marked in red in the manuscript.

5. The fourth paragraph in Introduction, “Nevertheless, …., these variants still rely on cascaded multi-stage CNNs. Therefore, we emphasize the design of particular good multi-scale features”. This sentence is confusing, I believe the propose method is still based on encoder-decoder architecture and used the cascaded multi-stage CNNs. Could the authors rephrase the sentence and highlight the difference between the proposed work and convention UNet?

Response: Thanks for the suggestion. We are very sorry that improper expression of this sentence makes you feel confused. In order to improve the quality of the manuscript, we have rephrased the sentence according to the reviewer’s comment as “U-Net and U-Net like models has been showing impressive potential in segmenting medical images, but the performance of these models will be poor when the target organ exhibits large shape and size variations among patients. Therefore, design good multi-scale features for medical images segmentation is essential. However, creating multi-scale representations requires feature extractors to use receptive fields of considerable variations to give a detailed account of parts, objects, or context at all possible scales. The natural way for CNNs to extract coarse-to-fine multi-scale features is to utilize a convolutional operator stack. Such inherent CNN capability of extracting multi-scale features leads to good representations for handling numerous medical image analysis tasks”. We hope this revision will satisfy you.

6. The legend of Figure 1 and Figure 2 are too small to see, could enlarge the legend.

Response: Thanks for pointing it out. We have replaced a new figure with enlarged the legend according to the reviewer’s suggestion. As shown in Fig1 and Fig 2.

7. The experiment setting separated the datasets into training and testing or cross-validation, since there is no validation set, how the final model is selected for testing?

Response: Thanks for the question. When the loss of the training model continuously fluctuates less than 0.01 within 20 epochs, the training is stopped, and then the last epoch is selected as the final model for testing. In order to make the manuscript more clear and easier for readers to understand, we have also made corresponding explanations in the last sentence of the second paragraph of the Experimental settings section, which is described as: “In addition, in this study we select the last epoch when the loss of the training model fluctuates less than 0.01 within 20 epochs as our final model for testing.”

8. Followed by the first concern, the authors highlighted the ability to “better learn and understand the features at different scales” at the Discussion and Conclusion, it would be more intuitive to visualize the attention maps at different scales for reader, and discuss the difference with non-multi-scale attention. Otherwise, I would suggest removing this statement.

Response: Thanks for your suggestion. In this paper, the Fig 9 is a visualization of the segmentation results for different medical images, in which lung CT images, retinal blood vessel images and blastocyst images have different resolutions, and different tissues also have different scales. It can be clearly seen from Fig 9 that the segmentation results of SA-Net are better than the traditional U-Net architecture. For example, whether it is a small-scale retinal blood vessel or a large-scale lung image, SA-Net shows better performance. Moreover, the Inner Cell Mass, Blastocoel, Trophectoderm, Zona Pellucida and Background have different scales in blastocyst image, but the SA-Net performance is still better compared with U-Net without multi-scale attention module. Therefore, these experimental results make us think that our proposed model can better learn the features at different scales. It may be that our expression is not clear enough, which makes you feel puzzled. We are very sorry about this, in response to this problem, we have revised this sentence in the manuscript and hope to make the expression of this sentence more clearly. The revised expression is shown in the red sentence in the Discussion and Conclusion section.

Summary:

This is a well-evaluated paper. The large-scale experiments and comparisons are impressive. The application on five different datasets is a bonus. Overall, this is a great study, it should be of potential interest of readers if authors could address above issues and further fine-tune some descriptions.

Reviewer #2:

General comments
Overall, this work proposes a scale-attention module to the Res2Net based UNet for the task of classification and segmentation in the medical imaging field. Experiments are done on 4 publicly available datasets and with one ablation study.

1) The claimed contribution is the proposed scale-attention module. However, according to the ablation study in Table V, the improvement of SA compared with Res2Net is really marginal, only 0.03% in ACC, 0.06% in AUC and 0.29% in F1, which could also be in the std range if running different algorithms for multiple times. Also, the analysis in the main text "In fact, the MCC, SE, ACC, AUC and F1 scores increase by 0.36%, 1.28%, 0.03%, 0.06%, and 0.35%, respectively" is not consistent with the numbers in Table V. Can authors check that?

Response: Many thanks for your suggestions and comments. We are very sorry that we didn’t explain clearly how the increase rate is calculated, which makes you confused. We calculate the increase percentage of MCC, SE, ACC, AUC and F1 as follows: , however, we know that you use the following formula to calculate the increase percentage: , so our results are not consistent with yours. If calculate increase percentage according to our formula, our result is correct. But it is true that our calculation method is prone to confusion, so we have revised the manuscript according to your calculation method. The revised version is consistent with your results.

In addition, comparing with Res2Net, indeed the improvement of SA is really marginal in terms of ACC, AUC and F1 indicators, but the improvement of sensitivity is significant, from 0.8148 to 0.8252 and the increase percentage is 1.04%. The sensitivity is a more important evaluation metrics compared to other metrics in medical image analysis. Therefore, the results of ablation experiments in Table 5 can demonstrate the importance of the proposed scale-attention module to a certain extent. In order to make the manuscript more clear and easier for readers to understand, we have revised the last few sentences of the Ablation study section as “When we added our proposed SA module, we can see that all indicators except for the specificity have increased, among which MCC, SE, ACC, AUC and F1 scores increase by 0.29%, 1.04%, 0.03%, 0.06%, and 0.29%, respectively. Particularly, the improvement of sensitivity is significant, these results can demonstrate the importance of the proposed scale-attention module to a certain extent.”

The revised sentence is:“When we added our proposed SA module, we can see that all indicators except for the specificity have increased, among which MCC, SE, ACC, AUC and F1 scores increase by 0.29%, 1.04%, 0.03%, 0.06%, and 0.29%, respectively. Particularly, the improvement of sensitivity is significant, so it can demonstrate the importance of the proposed scale-attention module to a certain extent.”

2) What is the computation cost comparison for the ablation study in Table V? I can image a marginal improvement can be obtained by a bit more computation cost.

Response: Thanks for the comments. We have added the computation cost in the last column of Table 5. As we can see from Table 5, it is indeed as you image that the computational cost is increasing with the improvement of performance. We are considering how to reduce the computational cost without reducing the performance of the model, this will also be one of our future research work.

3) Some important details of some datasets are missing. The training and testing split of the DRIVE dataset and the A/V classification dataset are missing. What is the GPU version you are using for the training and testing?
Response: Many thanks for your suggestions. a) For the first question, we have added the details of the training and testing split of the DRIVE dataset in the fourth sentence of the first paragraph of the Retinal vessel detection section, and also added the details of the training and testing split of the A/V classification dataset in the third sentence at the end of the first paragraph of the Artery/vein classification section. The description in the manuscript is “The DRIVE dataset consists of 20 training images and 20 testing images with the resolution of 584×565.”and “The training and testing subsets were obtained following the same schemes as those of the blood vessel detection problem, which means that 20 images are used for training and the remaining 20 are used for testing.”, respectively. We have marked red for these modifications in our manuscript. b) For the second question, the GPU version we are using for the training and testing is NVIDIA Tesla V100 GPU. We are very sorry that we missed this in the manuscript, we have added the GPU version in the Experimental settings section, as shown in the last sentence of this section.

4) In table IV, the reference number of Blast-Net is mixed to be the TernausNet paper.
Response: Thanks for pointing it out. We have changed the order of previous references [40] and [41] in the section of References. It should be noted that since we have added a new reference to the manuscript according to the reviewer1’s comments, the previous references [40] and [41] have become [41] and [42].

Pros:

1) The written is clear. The scale problem is a pain point in the medical imaging, even for the computer vision.

2) The experiments are done on many datasets.

Submitted filename: Response to Reviewers.docx

Click here for additional data file.

1 Feb 2021

PONE-D-20-30729R1

SA-Net: A Scale-Attention Network for Medical Image Segmentation

PLOS ONE

Dear Dr. Zhang,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we think the paper is acceptable after the final minor correction. Therefore, we invite you to submit a revised version of the manuscript that addresses the points soon to get your paper published.

Please submit your revised manuscript by Mar 18 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

We look forward to receiving your revised manuscript.

Kind regards,

Yuankai Huo, Ph.D.

Academic Editor

PLOS ONE

Additional Editor Comments (if provided):

This is paper is considered to be accepted after the final minor correction. Please make sure that the minor issues from the reviewer 1 is corrected.

"The emphasis of "improvement of sensitivity is significant" in the Ablation is not proper. First, the 1% improvement cannot be claimed as a significant one. Second, the comparison of sensitivity only is not fair since the specificity is worse. So I would suggest a rephrase. Apart from this, my other comments are addressed."

[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: Yes

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: (No Response)

**********

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: The authors did very well in addressing the comments. The revisions significantly clarify the concerns of the paper by including statements of their work. It is now clear that the author's approach is novel, and statements are clear.

Reviewer #2: The emphasis of "improvement of sensitivity is significant" in the Ablation is not proper. First, the 1% improvement cannot be claimed as a significant one. Second, the comparison of sensitivity only is not fair since the specificity is worse. So I would suggest a rephrase. Apart from this, my other comments are addressed.

All in all, although I still think the added SA structure based on Res2Net is a marginal improvement at the cost of extra parameters (params used from backbone to the backbone+SA: 34M >> 49M >> 161M >> 194M), this manuscript is clear-written with extensive valuable experiments, which would be good for readers to know. So I give an accept recommendation.

**********

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: Yucheng Tang

Reviewer #2: No

[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.

3 Feb 2021

Dear Editors and Reviewers:

We thank you very much for giving us an opportunity to revise our manuscript again. Thanks a lot for the helpful comments and recommends and thank you for your time spent. According to the reviewers' comments, we have revised the manuscript. Revised portion are marked in red in the paper.

The main corrections in the paper and the responds to the comments are as follows:

Reviewer #2: The emphasis of "improvement of sensitivity is significant" in the Ablation is not proper. First, the 1% improvement cannot be claimed as a significant one. Second, the comparison of sensitivity only is not fair since the specificity is worse. So I would suggest a rephrase. Apart from this, my other comments are addressed.

Response: Thanks for the suggestion. We think this comment is very reasonable and we have rephrased the sentence. In fact, the 1% improvement is slightly, so in order not to confuse the reader, we directly replaced the percentage of increase with values. And in manuscript we rephrase the sentence as:“When we added our proposed SA module, we can see that all indicators except for the specificity have increased, among which MCC, SE, ACC, AUC and F1 scores increase to 0.8055, 0.8252, 0.9569, 0.9822, and 0.8289, respectively. These results demonstrate the effectiveness of the proposed scale-attention module to a certain extent.”

Submitted filename: Response to Reviewers.docx

Click here for additional data file.

8 Feb 2021

SA-Net: A Scale-attention network for medical image segmentation

PONE-D-20-30729R2

Dear Dr. Zhang,

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,

Yuankai Huo, Ph.D.

Academic Editor

PLOS ONE

Additional Editor Comments (optional):

Reviewers' comments:

5 Apr 2021

PONE-D-20-30729R2

SA-Net: A Scale-attention network for medical image segmentation

Dear Dr. Zhang:

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. Yuankai Huo

Academic Editor

PLOS ONE