An attention-based weakly supervised framework for spitzoid melanocytic lesion diagnosis in whole slide images

Highlights：

• Spitzoid histological images are used for the first time to develop an automatic system able to predict the malignancy of spitzoid melanocytic.

• A new histology-based backbone based on feature refinement via attention layers and squeeze and excitation blocks to extract more accurate features is proposed.

• A novel framework based on inductive transfer learning to solve at the same time ROI selection and malignancy detection is developed.

• An end-to-end system able to classify melanocytic lesions by aggregating tumor regions' features is proposed.

• The heat maps extracted by computing the class activation maps are directly in line with the clinician's opinion since the highlighted regions of the histological images correspond to the interesting areas in which the experts focus for spitzoid diagnosis.