Accurate segmentation and analysis of 3D lesions in medical imaging is a critical challenge in computer-aided diagnosis, particularly when dealing with large volume of data. Glioma refers to a type of tumor that originates in the glial cells of the brain or spine and which is the most lethal disease worldwide presenting significant challenges for treatment. The primary challenges include accurately detecting the presence of brain tumors in patients and delineating the tumor boundaries. This paper proposes a fully autonomous glioma segmentation technique using a Deep Learning-based architecture Feature-Centered Multi-Mask network with Parallel Multi-Resolution Decoder (FCMM-PMRD), a novel framework designed to enhance segmentation accuracy through loss-guided progressive mask selection. This architecture efficiently captures multi-scale contextual information by integrating multiple parallel decoders operating at different resolutions. The N4 bias field correction algorithm is applied as a preprocessing step to address intensity non-uniformity in images and trained using categorical focal loss(CFL) as the loss function for finding the best feature map from the parallel decoder structure. The dynamic mask selection mechanism is guided by the loss function, ensuring adaptive feature refinement during training leads to the selection of best-decoded images. The decoded image is compared with the original mask for the minimum loss. The effectiveness of FCMM-PMRD is evaluated using publicly available 3D medical imaging datasets Brats-2020 & Brats-2023 and obtained superior segmentation accuracy and computational efficiency compared to existing state-of-the-art architectures. Our algorithm achieves Dice Similarity Coefficients (DSC) of 0.93, 0.93, and 0.89 for the whole tumor, enhancing tumor, and core regions.
M.C. et al. (Sat,) studied this question.