Application of Dataset Pruning and Dynamic Transfer Learning on Vision Transformers for MGMT Prediction on Brain MRI Images

Deep learning models, in particular vision transformer, are becoming more dependent on large and representative datasets. Their application on 3D medical images also has additional challenges. In this context, two primary solutions are used: either creating sub-volumes and then slicing them for patch embedding, or generating 2D slices from the 3D data before patch embedding. It is important to efficiently convert 3D images into 2D slices, removing less useful ones, in order to optimize computational resources. When analyzing brain MRI images, glioblastoma lesions present a significant challenge in identifying critical regions for MGMT promoter methylation, as these regions are typically small and difficult to detect. To address this challenge and improve the cost-effectiveness of 3D Brain MRI imaging, This paper propose a novel approach based on Dataset Pruning, leveraging Transfer Learning with Vision Transformers (MobileViTV2) and tracks prediction uncertainty across multiple epochs to evaluate and prune less informative images. The results show that high accuracy and precision can be achieved by minimizing noise and removing less relevant images, enabling the model to focus on the most critical data.