A Comparison of Autoencoders and Variational Autoencoders for Anomaly Detection in Dermoscopic Skin Lesion Imagery
Proceedings of the Canadian Conference on Artificial Intelligence · 19.05.2025
DOI: 10.21428/594757db.211a8674 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the Canadian Conference on Artificial Intelligence, online erschienen am 19.05.2025. Verfasst von Errol DaRocha und 2 weiteren Autor:innen.
Abstract (Originalsprache)
Early detection and diagnosis of skin abnormalities are crucial for effective treatment and management of skin diseases. This paper explores the application of deep learning techniques for skin tissue analysis, focusing on the detection of abnormalities from dermoscopic images. Unsupervised learning methods such as Autoencoders (AE) and Variational Autoencoders (VAE) save resources by eliminating the need for labeled data, making them more efficient and scalable than supervised learning. We compare the performance of AE and VAE architectures in developing a robust model capable of distinguishing between benign and malignant skin lesions. This study uses the MNIST HAM10000 dataset of 10,015 dermoscopic images, divided into seven classes that represent benign and malignant diagnostic categories. The data set was divided into benign (normal) and malignant (anomalous) cases. The models were trained to learn features of the normal data and generate reconstructions of these images. An optimal decision boundary is chosen to classify images as benign or malignant based purely on their reconstruction error. Experimental results, averaged over 30 training runs, demonstrate that the AE outperforms the VAE in accuracy (71.84/% vs. 67.79/%), F1-score (73.03/% vs. 67.33/%), and FNR (23.76/% vs. 33.58/%). An evaluation of Autoencoder reconstructions using Shapley Additive Explanations (SHAP) analysis was performed to visualize and identify the pixel-wise attributions of anomalies thereby providing insight as to which factors most influenced model reconstruction and, therefore, decisions in label assignment for each AE approach. Some general patterns were clear, particularly regarding high attribution assignments to centralized pixel regions of the most anomalous images. These findings suggest that the AE architecture is promising for automated skin cancer detection, which can lead to more accurate and timely diagnoses and improved patient outcomes.
Zitieren
APA
DaRocha, E., McIntyre, A., & Bouzar-Benlabiod, L. (2025). A Comparison of Autoencoders and Variational Autoencoders for Anomaly Detection in Dermoscopic Skin Lesion Imagery. Proceedings of the Canadian Conference on Artificial Intelligence. https://doi.org/10.21428/594757db.211a8674
Vancouver
DaRocha E, McIntyre A, Bouzar-Benlabiod L. A Comparison of Autoencoders and Variational Autoencoders for Anomaly Detection in Dermoscopic Skin Lesion Imagery. Proceedings of the Canadian Conference on Artificial Intelligence. 2025. doi:10.21428/594757db.211a8674
BibTeX
@article{DaRocha2025,
author = {DaRocha, Errol and McIntyre, Andrew and Bouzar-Benlabiod, Lydia},
title = {A Comparison of Autoencoders and Variational Autoencoders for Anomaly Detection in Dermoscopic Skin Lesion Imagery},
journal = {Proceedings of the Canadian Conference on Artificial Intelligence},
year = {2025},
doi = {10.21428/594757db.211a8674}
}
EUVIMED