International Journal of Science, Engineering and TechnologyAuthors: Sharmila Arun Chopade, Dr. Akash Saxena
Abstract: Liver tumor segmentation of contrast-enhanced CT (computerized tomography) scans is an important step for computer-aided liver cancer diagnosis, treatment, and follow-up. Although deep learning approaches have recently been significantly advancing the task of medical image segmentation, liver tumor segmentation still faces many challenges since liver tumors usually have non-regular shapes, fuzzy boundaries, and heterogeneous enhancement patterns. Besides, class imbalance between lesions and background tissues is also very high. Existing models mostly predict a final liver tumor mask without providing any kind of information on the reliability of this prediction. The absence of uncertainty quantification is one of the most serious problems of medical imaging due to the necessity of reviewing uncertain regions by experts.In this work, we propose UQ-TSE-Net (uncertainty-quantifying total segment enhancement network), a novel uncertainty-aware lightweight architecture for reliable liver tumor segmentation of contrast-enhanced CT scans. The proposed approach combines efficient feature extraction via convolutions, attention-based feature refinement, boundary awareness, and uncertainty estimation using Monte Carlo dropout. The first two blocks form a main segmentation branch, while the latter allows improving the performance of segmentation in ambiguous regions. Finally, the uncertainty of the model predictions is estimated using the Monte Carlo dropout approach that produces a set of different predictions. Based on these predictions, a pixel-level uncertainty map is obtained highlighting low-confidence regions of liver tumors, especially those located at fuzzy tumor margins.For the experiments, we will use the Liver Tumor Segmentation Challenge dataset (LiTS17). The comparison will be made to widely-used segmentation models such as UNet, Attention UNet, UNet++, DeepLabV3+, TransUNet, UNETR, and nnU-Net. The models' performance will be evaluated using classical segmentation metrics like Dice score, Intersection over Union, precision, and recall and additional reliability-oriented metrics, like Expected Calibration Error and Brier score.UQ-TSE-Net will allow getting not only an accurate tumor mask but also uncertainty information that is crucial for further interpreting the results. In the current study, we make another step towards obtaining reliable computer-aided liver tumor analysis with the help of uncertainty-aware lightweight architecture.
DOI: https://zenodo.org/records/22974329
