Enhancement of Classification Models using Outlier Detection Algorithms: Melanoma Diagnosis Case Study

Abstract

Researchers have applied different machine learning and deep learning models to attempt to diagnose skin cancer, or melanoma. Because of the nature of the lesions, they were partially successful in reaching limited accuracies. Some of the melanoma lesions were still misclassified as benign, which causes late treatment in this deadly disease. In this study, we proposed an alternative method to improve the model's vulnerability in real-world examples. Using the model, we first identify benign lesions, most of which are truly benign, and very few are actually misclassified melanoma. Benign labeled test data are then injected into the benign training dataset for verification. Since misclassified melanoma lesions among those will diverge from the majority of the benign lesions, they fall into the outlier category. One-Class SVM (OCSVM), Isolation Forest (IF), and Local Outlier Factor (LOF) were used for outlier detection using an additional effective features obtained from lesion image segmentation. Accuracy, recall, specificity, precision, F1-skor and ROC-AUC performance metrics have been obtained for the evaluation of the models. The majority of these misclassified melanoma lesions are identified as outliers, increasing melanoma diagnosis accuracy from 94% up to 98%.