AUTHOR=Yang Xiaoguo , Zheng Yanyan , Mei Chenyang , Jiang Gaoqiang , Tian Bihan , Wang Lei 

TITLE=UGLS: an uncertainty guided deep learning strategy for accurate image segmentation

JOURNAL=Frontiers in Physiology

VOLUME=Volume 15 - 2024

YEAR=2024

URL=https://www.frontiersin.org/journals/physiology/articles/10.3389/fphys.2024.1362386

DOI=10.3389/fphys.2024.1362386

ISSN=1664-042X

ABSTRACT=We are in receipt of your decision in regard to our manuscript, entitled "UGLS: An uncertainty guided deep learning strategy for accurate image segmentation" (Manuscript ID: 1362386). We thank the Associate Editor and the referees for their helpful comments and revised the manuscript as recommended to improve its quality. Our responses to each of the comments made by the Editor and the referees follow:The paper proposes an uncertainty guided deep learning strategy (UGLS) to improve image segmentation accuracy. It first trains the U-Net to obtain coarse segmentation. Boundary uncertainty maps are generated from coarse results and used along with original images for fine segmentation. My major concerns are listed below: 1. No comparison with very recent and top-performing methods like Tranformer-based models. Compare with state-of-the-art methods like Vision Transformers for segmentation. ANSWER: As suggested, we provided additional experiments based on Swin-Unet and TransUNet for performance comparison. Thank you! 2. Details of implementations and hyperparameter tuning are lacking. Provide implementation details of network architecture, hyperparameters, data augmentation etc. ANSWER: Now, we provided implementation details of the U-Net and data augmentation in the revision (see Section 2.6). Thank you! 3. Quantify and analyze uncertainty in segmentation predictions. Also, investigate different approaches to generate boundary uncertainty maps from coarse results ANSWER: Thanks for your comments! (1) We summarized the effect of boundary uncertainty maps in our developed method based on three different use strategies (see Table 6). These uncertainty maps were mainly used to improve segmentation performance by identifying a potential boundary region of each object and excluding the influence of irrelevant background. (2)There are many different approaches to generate boundary uncertainty maps, but they had a small effect on the final segmentation performance.4. In the introduction, cite relevant review papers in uncertainty quantification (e.g.