MSDTrA: A Boosting Based‐Transfer Learning Approach for Class Imbalanced Skin Lesion Dataset for Melanoma Detection

Pigmented skin lesion datasets comprise a higher percentage of benign lesion than the malignant lesions which lead to the class skewness issue in the dataset. Classifiers trained for analyzing the automated dermatoscopic pigmented lesions often suffer from data scarcity. Transfer learning permits to leverage the knowledge from the source domain to train a classifier towards the target domain when the data is rare. Importing knowledge from multiple or several sources towards increasing the chance of searching a source closer to a target may alleviate the negative transfer. A framework is proposed in this work to transfer knowledge from multiple different sources utilizing AdaBoost, TrAdaBoost and MultiSource Dynamic TrAdaBoost (MSDTrA), for melanoma detection. The effectiveness of the proposed framework is evaluated on four benchmark skin lesion datasets namely, PH2, ISIC16, ISIC17, and HAM1000 which demonstrate promising performance by alleviating negative-transfer by increasing multiple different sources.