Contrastive Self-Supervised Ensemble Transfer Learning for Robust Skin Cancer Classification and Early Detection

Melanoma is one of the most dreaded types of cancer in the world today, and therefore, early detection becomes crucial. Deep learning models need annotated training data, and such training data are difficult and expensive to acquire. To solve this challenge, we introduce a new framework called the Contrastive Self-Supervised Ensemble Transfer Learning (CSSL-ETL) that combines the techniques of CSSL and ETL to improve the feature learning and classification accuracy of the model.

The CSSL-ETL framework integrates Contrastive Self-Supervised Learning (CSSL) and Ensemble Transfer Learning (ETL) techniques. Utilises CSSL, which pre-trains models on a vast array of images and enhances the generalization of models to unlabeled skin images, whereas ETL captures various Feature extraction power evolves the ConvNeXt-Large, Swin Transformer V2, and EfficientNetV2 models.

Using the same metrics on both datasets, ISIC and HAM10000, such accuracies are 94.6%, precisions of 93.8% and recalls of 91.5%, as well as the AUC-ROC of 96.1%, which is higher than the ResNet-50, EfficientNetV2, and Swin Transformer feeds forward neural network models. The assessment of the confusion matrix also reveals low misclassifications, particularly in the ability to identify melanoma. Coping with clinical and thermoscopic images in a combined manner increases the diagnostic capabilities of the system. On the same note, federated learning takes into consideration the private architecture of the model across institutions in the context of AI. The incorporation of Grad-CAM++ and the Bayesian estimate of uncertainty enhances the models' transparency and, ultimately, the clinicians' confidence in those models.

Compared to the previous methods, CSSL-ETL represents the features better and strengthens the classification ability and generalization ability. As for future work, real-time m-health applications as well as data fusion using multiple sources of data, which will enhance the automation of skin cancer detection, will be the next areas of concern.