Sub-region-based patching for improved sub-region-based brain tumor segmentation in 3D MRI scans
Abstract
Early diagnosis of brain tumors is imperative to save lives. However, automated diagnosis faces several challenges, such as the use of 3D MRI scans, which demand substantial computational resources especially since medical images are highly prone to class imbalance, making models biased toward the majority class. Various patching techniques, such as fixed, random, and random biased, have been employed to address these issues. However, they often fail to effectively preserve the entire tumor due to cropped tumor regions. Conversely, overlapping patches that preserve tumor regions exacerbate class imbalance because of redundant background pixels. Furthermore, the number of patches increases as the overlap decreases, resulting in a nonlinear relationship that introduces redundancy and affects overall segmentation performance. Moreover, these techniques do not enable the model to focus on regions of interest corresponding to specific tumor subregions.In this study, we propose a novel subregion based patching technique that selects patches from both the whole tumor region (WT Patch) and the core tumor region (TC Patch) without relying on ground-truth information, thereby making it suitable for inference in segmentation tasks. This patching methodology centers the tumor within each patch, effectively preserving its overall morphology and preventing cropped tumor boundaries. The proposed approach enhances model training by enabling independent learning from distinct subregions. We evaluate our method against state-of-the-art patching methodologies and conduct extensive experiments using the BraTS 2020 dataset. Our findings demonstrate substantial improvements, achieving Dice scores of 96\% for the whole tumor, 90\% for the core tumor, and 92\% for the enhancing tumor region, outperforming existing state-of-the-art methods.