Brain Tumour Segmentation and Edge Detection Using Self-Supervised Learning

Brain tumour segmentation is a critical in medical imageanalysis, facilitating diagnosis and treatment planning in neurosurgery.Brain tumour segmentation with supervised learning show robust resultsin medical imaging, however, it requires sufficient amount of annotateddata for effective learning. It is important to detect boundaries of tumoursubregions accurately in fine-grained segmentation. We propose a novelapproach that uses a unique dual-decoder architecture, focusing on edgeidentification and segmentation accuracy enhancement. Utilizing a dualdecoder 3D-UNet model, we prioritize accuracy and fine-grained detailsin tumour segmentation and introduce an additional tumour edgedetection task, aiming to move beyond traditional single-decoderapproaches. Incorporating a 3D SimSiam network as the self-supervisedpretraining technique, we aim to address the limitation of annotated dataand enhance the segmentation accuracy. Our model surpasses manysupervised variants of U-net architectures and self-supervisedapproaches, highlighting the importance of edge detection in tumoursegmentation. The proposed approach enhances segmentation accuracyby showing an accuracy of 98.1% and provides critical boundary detailsfor clinical decision-making. Visualizations of segmentation and edgemasks further validate the effectiveness of the proposed method.