Human knowledge-based artificial intelligence methods for skin cancer management: Accuracy and interpretability study

Abstract

Skin cancer management, including monitoring and excision, involves sophisticated decisions reliant on several interdependent factors. This complexity leads to a scarcity of data useful for skin cancer management. Deep learning achieved massive success in computer vision due to its ability to extract representative features from images. However, deep learning methods require large amounts of data to develop accurate models, whereas machine learning methods perform well with small datasets. In this work, we aim to compare the accuracy and interpretability of skin cancer management prediction 1) using deep learning and machine learning methods and 2) utilizing various inputs including clinical images, dermoscopic images, and lesion clinical tabular features created by experts to represent lesion characteristics. We implemented two approaches, a deep learning pipeline for feature extraction and classification trained on different input modalities including images and lesion clinical features. The second approach uses lesion clinical features to train machine learning classifiers. The results show that the machine learning approach trained on clinical features achieves higher accuracy (0.80) and higher area under the curve (0.92) compared to the deep learning pipeline trained on skin images and lesion clinical features which achieves an accuracy of 0.66 and area under the curve of 0.74. Additionally, the machine learning approach provides more informative and understandable interpretations of the results. This work emphasizes the significance of utilizing human knowledge in developing precise and transparent predictive models. In addition, our findings highlight the potential of machine learning methods in predicting lesion management in situation where the data size is insufficient to leverage deep learning capabilities.