Learning a Meta-Ensemble technique for skin lesion classification and novel class detection

Abstract

The frequency and fatality rates associated with skin Melanoma requires an accurate and efficient detection methodology to enable early medical diagnosis. Artificial Intelligence (AI) augmented detection methods aim at achieving this goal while reducing the costs and time involved in traditional methods. This work utilizes a two-level ensemble learning technique (trained with weighted losses) to improve accuracy over individual classification models. The ensemble technique alleviates over-fitting due to class imbalance in the dataset, achieving a Balanced Multi-class Accuracy (BMA) score of 0.591 without unknown class detection. The algorithm was extended by appending the proposed CS-KSU module collection to detect the presence of images belonging to novel classes during test time. The extended algorithm secured an Area Under the ROC Curve (AUC) score of 0.544 for the unknown class. Our algorithm’s performance is at par with the current state-of-theart for this task1 .