Unsupervised Clustering to Reduce Overfitting Issues in Ensemble Deep Learning Models for Skin Lesion Classifications

Abstract

Class imbalance in skin lesion cancer is pronounced, which in turn impacts the performance of deep learning models for classification tasks. In the case of using dermoscopic images, training a convolutional neural network-based ensemble deep learning model on 7 classes from the International Skin Imaging Collaboration (ISIC) dataset 2018 with the 2 minor classes as malignant yields decent performance in classifying skin lesions. However, the predictive ability of the model on unseen data, ISIC 2019 is unsatisfactory. To narrow the gap of overall accuracy between the training and testing tasks, Part 1 of this study conducts 2 different sampling techniques: augmentation and clustering algorithms. The results of this study—Part 2, show that 4 samples taken from Self-Organizing Maps, Hierarchical and K-Means techniques have narrower gaps of the overall accuracies between the training and testing sets. The best-balanced results of the overall accuracies for the training and testing sets are 82 and 43%, respectively, if one cluster of melanocytic nevi from hierarchical clustering is used, while augmentation gives 92 and 40%, respectively.