This study uses the MURA dataset to evaluate various transfer learning models for detecting anomalies in the humerus bone. Six pre-trained convolutional neural networks—EfficientNetB0, ResNet-50, InceptionV3, MobileNetV2, VGG-19, and DenseNet—were employed to perform the classification task on 1560 X-ray images. Given the significant public health impact of musculoskeletal conditions, there is a pressing need for accurate and prompt diagnosis to mitigate prolonged recovery times, increased healthcare costs, and potential disability. This research leverages the strengths of deep learning, particularly transfer learning, to address the challenges posed by limited, high-quality medical image datasets. The DenseNet model gave an accuracy of 82%, and the EfficientNetB0 model gave an accuracy of 81%, outperforming other models. The results indicate that combining features from different CNN models can improve the detection accuracy of humerus abnormalities. However, due to the black-box nature of these models, additional efforts are needed to eliminate their decision-making processes, with the goal of increasing transparency and trust in deep learning-based diagnostic tools.