Determining progress during wound healing is crucial for effective diagnosis and treatment. Previous works have solved this task using methods paying attention to specific regions of the image. However, we explore an alternative, nonlocal attention approach and implement a cross-layer attention mechanism that focuses on the areas of interest and considers related spatial regions of the wound. Experimental results and visual representations show that adding cross-layer modules to mid-level and top-level layers enables better classification of wound healing stage and generalization.

Google Colab link for the code can be accessed here.

Google Drive link for the dataset and weights is here.