LAUSR

BACKGROUND The distinction between in situ (MIS) or invasive melanoma is challenging even for expert dermatologists. The use of pretrained convolutional neural networks (CNNs) as ancillary decision systems needs further research. OBJECTIVE To develop, validate and compare three deep transfer learning algorithms to predict between MIS or invasive melanoma and < or ≥ 0.8 millimetres of Breslow thickness (BT). METHODS A dataset of 1,315 dermoscopic images of histopathologically confirmed melanomas was created from Virgen del Rocio University Hospital and open repositories of the ISIC archive and Polesie et al. The images were labelled as MIS or invasive melanoma and < or ≥ 0.8 millimetres of BT. We conducted three trainings, and overall means for ROC curves, sensitivity, specificity, positive and negative predictive value, and balanced diagnostic accuracy outcomes were evaluated on the test set with ResNetV2, EfficientNetB6, and InceptionV3. The results of ten dermatologists were compared with the algorithms. Grad-CAM gradient maps were generated, highlighting relevant areas considered by the CNNs within the images. RESULTS EfficientNetB6 achieved the highest diagnostic accuracy for the comparison between MIS and invasive melanoma, and < 0.8 versus ≥ 0.8 of BT were 61% and 75%, respectively. For the latter, ResNetV2, with an area under the ROC curve of 0.76, and EfficientNetB6, of 0.79, outperformed the results obtained by the dermatologist group with 0.70. CONCLUSIONS EfficientNetB6 recorded the best prediction results, overcoming dermatologists for the comparison of 0.8 mm of BT. DTL could be an ancillary aid to support dermatologists' decision in the near future.