XAI-ResUNet: Analysing the Impact of Pre-training in ResUNet Architectures for Multiple Sclerosis Lesion Segmentation using EigenGradCAM
Vayangi Ganepola, Prateek Mathur, Oluwabukola Adegboro, Julia Dietlmeier, Aonghus Lawlor, Noel E. O’Connor, Claudia Mazo
Code Available — Be the first to reproduce this paper.
ReproduceCode
- github.com/VishmiVishara/XAI-ResUNetOfficialIn paperpytorch★ 1
Abstract
Multiple Sclerosis (MS) is a chronic disease that causes lesions in the central nervous system. Diagnosing MS is a challenging process that strongly relies on Magnetic Resonance Imaging (MRI) and exhibits high and largely unexplained variability across patients. In this paper, we propose a novel XAI-ResUNet architecture by integrating ResNet-50 as an encoder in a U-Net-like architecture and demystifying the detection and segmentation process by incorporating EigenGradCAM into the encoder. We used the training and testing sets of the MSSEG-2016 dataset for our experiments. We compared three architectures with ResNet-50 pre-trained on ImageNet, RadImageNet, and without pre-training. Our preliminary results show that ResNet-50 pre-trained on ImageNet outperforms the others, achieving a DSC of 0.638. We also demonstrate how pre-training on large image datasets can affect MS lesion detection and segmentation. Our code is publicly available at https://github.com/VishmiVishara/XAI-ResUNet.