Combining Weakly Supervised Segmentation with Multitask Learning for Improved 3D MRI Brain Tumour Classification.

Segmentation plays a crucial role in computer-aided medical image diagnosis, as it enables the models to focus on the region of interest (ROI) and improve classification performance. However, medical image datasets often lack segmentation masks due to cost. In this work, we have introduced a novel end-to-end pipeline that enhances classification by incorporating ROIs generated by a weakly supervised method as an auxiliary task in a deep multitask learning framework. We have demonstrated that our approach outperforms conventional methods that only rely on classification. By using weakly supervised segmentation, we are able to leverage pixel-wise information, which ultimately leads to improved classification performance.