Fig. 4.

Examples of correct predictions highlighted by Gradient-weighted Class Activation Mappings (Grad-CAMs) for multilabel classifications. A and D are the original images, showing a fracture at the humeral head and the humeral shaft with osteosynthesis material (A) and a dislocated shoulder endoprosthesis (D). In B, the model showed a strong activation in image regions of the fracture and in C a strong activation for image regions depicting osteosynthesis material, which lead to the models final classification result. E shows activation of the model in the area where the humeral head should be, leading to the recognition of the dislocation, and F shows the activation for detection of the endoprosthesis