Publication

Abstract : Biomedical image segmentation is an important task in medical diagnosis, treatment planning and disease monitoring and current deep learning models are often limited by their inability to handle structural distortions as well as their weak boundary delineation and low robustness in varying imaging conditions. To overcome the difficulties, this study proposes a deep morphological learning framework for biomedical segmentation named BioVisionNet to improve the shape preservation and boundary accuracy of the segmentation results. The goal is to combine the learnable morphological operator, multi-scale feature encoding and attention-based decoding in order to overcome the limitations of traditional CNN and transformer models. The proposed methodology takes into account a Morphological Feature Encoding Layer, multi-path BioVision encoder and boundary refining attention decoder that is trained with a structural consistency loss function. Experiments are performed on BRATS 2023 dataset and it achieves a better performance with Dice Score of 0.953, IoU of 0.912, and Hausdorff Distance of 4.8 which is better than 5 stateof-the-art methods. The findings attest to the fact that BioVisionNet brings much better structural fidelity, segmentation reliability and clinical applicability. In conclusion, in this work, a solid basis for a morphology-aided segmentation has been built and holds great promise to be applied to actual medical imaging workflows in the future.