Class-Agnostic Feature-Learning-Based Deep-Learning Model for Robust Melanoma Prediction

Abstract

Skin lesion images are relatively easy to obtain, but variations in imaging conditions, particularly lesion positioning, can cause inconsistencies in deep learning model activations, impacting diagnostic reliability. In this study, a robust deep-learning model for melanoma prediction was developed using class-agnostic activation maps (CAAMs) for enhanced diagnostic accuracy and reliability by addressing issues related to image variability and transformation robustness. The international skin imaging collaboration (ISIC) 2017 and 2019 datasets, focusing on melanoma and nevus, were used. The performance was evaluated using the area under the receiver operating characteristic curve (AUROC), while Dice scores were used for robustness evaluation. The proposed model achieved an AUROC of 0.954 for ConvNeXt on the ISIC 2019 dataset, with Dice scores of 0.664 and 0.457 for ConvNeXt and ResNet, respectively. For the ISIC 2017 dataset, the proposed model achieved an AUROC of 0.843 for ConvNeXt, with Dice scores of 0.557 and 0.306 for ConvNeXt and ResNet, respectively. The proposed CAAM-based method improved the melanoma prediction and lesion recognition accuracy by ensuring robust and consistent CAAMs.