PB-FELTuCS: Patch-Based Filtering for Enhanced Liver Tumor Classification and Segmentation

Bheeshm Sharma, P. Balamurugan

Abstract: In this paper, we introduce PB-FELTuCS, a hierarchical deep neural network architecture which first performs enhanced 3D liver segmentation and subsequently employs a 3D patch-based filtering algorithm over the 3D liver segments to achieve 3D liver tumor classification and segmentation on LiTS dataset [1]. Despite the simplicity of our liver segmentation network, it surpasses recent benchmarks by achieving an impressive Dice score of 0.98, facilitated by our proposed weighted version of Exponential Logarithmic Dice (ELDice) loss [20]. Furthermore, we propose a filtering approach to extract meaningful 3D patches from the segmented liver, which are then used (as opposed to full volumes) during training of our tumor classification and segmentation networks. This approach enables simpler networks to obtain an accuracy of 89.1% in tumor classification and a 0.747 Dice score for tumor segmentation, highlighting the importance of effective training strategies as an alternative to complex neural network architectures, in enhancing the precision of volumetric medical assessments. Code is available at https://github.com/BheeshmSharma/PBFELTuCS_MICAD2023.