Musculoskeletal Disorders (MSDs) affect 1.7 billion people worldwide, causing pain and mobility issues [22]. Though convolutional neural network (CNN) models can detect MSDs, their opaque nature raise concerns regarding transparency and accountability. The class activation map (CAM) method provides insight into CNN predictions by visually highlighting regions of interest or local discrimination. This assists medical professionals in visualizing the decisions made by the CNN model. CAM utilizes activation maps of a specific convolutional layer, typically the last convolutional layer, along with the predicted class score. CAM’s local discrimination effectiveness depends on activation maps and weight computation. Therefore, CAM’s ability to local discrimination depends on the dataset and the CNN model. Therefore, it is essential to conduct an empirical study on datasets with popular CNN models used in MSD classification. This is important for ensuring unbiased decision-making in medical imaging. This study quantitatively evaluates various CAMs, including CAM, Grad-CAM, Extended-CAM, Smooth Grad-CAM and Score-CAM. Three CNN models, ResNet-50, DenseNet-169, and Inception-v3, were trained on the MURA dataset. The evaluation employs quantitative measures such as average drop, average increase, maximum coherency, and minimum complexity. Average drop coherency complexity (ADCC) is a single metric that compares qualitative measures. Results indicate that Smooth Grad-CAM and Score-CAM outperform in local discrimination regions for predictions. Smooth Grad-CAM outperforms in average drop, average increase, and ADCC for ResNet-50. However, for DenseNet-169 and Inception-v3, Score-CAM excels in ADCC.