Fully Automatic Thoracic Cavity Segmentation in Dynamic Contrast Enhanced Breast MRI Using Deep Convolutional Neural Networks

Round 1

Reviewer 1 Report

Fig 3 after mask application is not clear

Hyperparameters chosen can be explained clearly.

Refinement algorithms applied is mentioned but it is not explained clearly

Minor corrections few sentences are very lengthy and its difficult to understand

Author Response

We thank the reviewer for their constructive feedback. We have made the following changes to the manuscript to address the concerns raised:

• We have added a sentence in the caption of Figure 3 to explain the effect of the mask application, as well as adding a red circle to direct the attention of the reader to the change.
• We have added details on training hyperparameters (section 3.5.2)
• We have added further information about the refinement algorithm (line 369-371)
• We have reduced the average length of sentences

Reviewer 2 Report

The paper addresses a significant issue in medical imaging: the segmentation of the thoracic cavity in dynamic contrast-enhanced breast MRI. The manuscript is well-written, with the introduction effectively emphasizing the importance of Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) in breast cancer diagnosis and the challenges involved in distinguishing tissue changes induced by contrast agents from those occurring in internal organs. The paper clearly outlines its primary objective: the development of a data-efficient approach for thoracic cavity segmentation. In the methodology section, a comprehensive description is provided, particularly regarding the utilization of UNet-like architectures based on ResNet. They also evaluate the various configurations of ResNet model. The results section claims that the proposed methodology outperforms the current state-of-the-art in terms of data efficiency and similarity index when compared to manually segmented data. The authors have used a primary dataset for training, testing, and validation of the deep learning model. Additionally, they have augmented their dataset with synthetic data, thereby increasing the volume of data available for model training and demonstrating a thorough approach to data management.

Overall, the paper is well-structured and effectively conveys the author's experiments and findings.

Author Response

We thank the reviewer for the kind words.

Reviewer 3 Report

 I think the authors should add more images to make readers understand the data. It took time for me to understand the data. In addition, the authors should add the visual results for the best and worst methods as seen in Figure 1. It should be better to see the visualt results, e.g. the automatic segmented area vs the manual segmented area.

Except that, it is a good study and the results are sufficient. My decision is minor revision for the paper.

Author Response

We thank the reviewer for the constructive feedback. We have added Figure 7 to the manuscript which aims at showcasing the differences between the worst performing architecture and the best performing one.