Forward diffusion guided reconstruction as a multi-modal multi-task learning scheme

Multi-Modal MRI images offer various perspectives for identifying regions of interest in the brain. Previous studies have successfully utilized Deep Learning methods in tasks, such as, segmentation and classification of MRI images, but the proper utilization and integration of Multi-Modality is still an open area of study. Some studies use Multi-Task training scheme which utilizes an auxiliary task like reconstruction for better performance. This paper proposes a novel Multi-Task learning scheme that utilizes different modalities of MRI images to improve brain region segmentation and classification, where the forward diffusion process and a time projection module is used to incorporate a guided reconstruction task. Our experimental results show that the proposed Multi-Task learning strategy outperforms the vanilla Single-Task training scheme by 2.4% in segmentation and 2.7% in classification tasks.