Abstract
Meniscus tears are among the most common injuries of the knee joint and require accurate and timely diagnosis. Deep learning–based segmentation models are widely used to automate the analysis of medical images; however, the limited interpretability of their predictions necessitates the application of explainable artificial intelligence (XAI) methods. In this study, the Grad-CAM and RISE methods are compared for the explainable interpretation of meniscus tear segmentation in magnetic resonance imaging (MRI) using the YOLOv8m-seg model. The study was conducted on a clinical dataset comprising 4000 2D MRI slices (normal/tear) with manually annotated segmentation masks. The evaluation of the methods is based on the Insertion and Deletion metrics, as well as on visual analysis of the explanation maps on the test set. The results demonstrate the superiority of the RISE method in generating more localized explanations supported by quantitative evaluation, thereby confirming the potential of model-agnostic XAI approaches for medical image analysis.