Tracking Lesion Evolution Using a Boundary Enhanced Approach for MS Change Segmentation (BEAMS)

Prateek Mathur, Brendan S. Kelly, Ronan P. Killeen, Aonghus Lawlor

Abstract: Multiple sclerosis (MS) is a chronic disease of the Central Nervous System (CNS) primarily characterised on Magnetic Resonance Imaging (MRI) by hyper-intense lesions. Accurate and timely identification of new lesions or growth in existing lesions has a tremendous impact on treatment planning and patient care. Early detection can lead to prompt intervention possibly slowing the disease progression. However, this task is challenging, costly and often prone to observer bias. Our approach consists of three key components (1) redefining the lesion segmentation problem as a change detection challenge focusing only on new and enhancing lesions; (2) adding an additional boundary detection task to an existing Convolutional Neural Network (CNN) to provide detailed context regarding the shape, size, and evolution of enhancing lesions; (3) evaluating the \(F_{\beta }\) loss objective as an alternative to traditional segmentation loss objectives, aiming to optimise the model’s sensitivity and specificity. Furthermore, we employ Gradient-weighted Class Activation Mapping (GradCAM) to visualise the network’s attention, aiding in the interpretation and validation of our approach. By providing an objective evaluation, our method aims to streamline clinical decision-making processes for improved patient outcomes.

External IDs:dblp:conf/miccai/MathurKKL24