Deep learning for segmentation of ischemic stroke lesions from MRI scans

Abstract

Convolutional Neural Networks (CNN) such as U-Net have been extensively used for medical image segmentation tasks. However, CNNs have limitations in learning global feature representations in images due to their local receptive fields. This prompted researchers to look at other methods to improve the modelling of global dependencies. In recent years, transformer-like architectures with self-attention mechanisms for modelling long-range dependencies have shown better performance than CNN-based architectures in computer vision applications.

In this project, deep learning techniques involving transformer-like architectures were explored for ischemic stroke lesion segmentation from Magnetic Resonance Imaging (MRI) scans. Specifically, this study focused on using the hybrid CNN-Transformer network, UNet Transformers (UNETR), to compare with pure CNN architectures. Furthermore, to address the issue of transformers needing large amounts of labelled data to perform well, the self-supervised learning method, Self-Distillation with No Labels (DINO), was utilised for network pre-training. In order for the DINO approach to work on the 3D-based UNETR architecture, we have modified the method to accept 3D images as input.

We demonstrated that with the pre-trained DINO weights, the ischemic stroke lesion segmentation performance of UNETR improved when supervised learning was performed on the Anatomical Tracings of Lesions After Stroke - Release 1.2 (ATLASR1.2) dataset.