Integrating the prior shape knowledge into deep model and feature fusion for topologically effective brain tumor segmentation

Effective segmentation of brain tumor from MRI scans is crucial for clinical diagnosis. However, deep learning models prioritize either local or global features, failing to adequately capture topological features or integrate prior shape knowledge. Deep learning techniques totally rely on the loss function optimization and due to the lack of explicit form of prior knowledge, they may struggle to generate the accurate tumor shapes. This often leads to significant errors in capturing the underlying shapes or sub structures of tumor. These limitations lead to challenges such as topological errors, patchy patterns, and unrealistic structures in segmentation. To address these shortcomings, we proposed a novel technique termed TDAConvAttentionNet that captures the local, global and topological features and integrates the prior shape, captured using persistent homology, into the deep model to enhance segmentation results and reduce topological errors. By incorporating prior shape information, our proposed model guides the segmentation process, resulting in more precise delineation of tumor regions and mitigating the effect of fragmented structures or patchy patterns. The proposed technique leverages the strength of ConvNeXt to capture the local features and multi-headed self-attention to capture the global features without using the entire transformer encoder to keep the model less complex. Evaluation on the publicly available BraTS2021 dataset demonstrates the effectiveness of our approach, assessed through dice score, IOU score, and Jaccard distance metrics. The proposed approach achieved dice scores of 89.36%, 87.36%, and 89.98% on the validation dataset for Whole Tumor, Tumor Core, and Enhancing Tumor, respectively. On the test dataset, the method achieved Dice scores of 87.74%, 85.86%, and 86.62% for Whole Tumor, Tumor Core, and Enhancing Tumor respectively. Our results highlight better segmentation performance compared to the existing state-of-the-art methods.