Title:
AI Algorithms for Skin Cancer Detection: From Convolutional Networks to Vision Transformers
Abstract:
Artificial intelligence has shown substantial promise for improving the accuracy, consistency, and scalability of skin cancer screening. This talk presents our recent work on deep learning algorithms for automated skin lesion analysis, with a particular focus on dermoscopic images. We will discuss the development of an ensemble-based classification framework that achieved third place in the ISIC MILK10K Challenge, illustrating how model architecture, data augmentation, training strategy, and model fusion can improve performance on large and heterogeneous dermatologic image datasets.
We will also present our work on vision transformer (ViT)-based melanoma classification, published in Cancers. Unlike conventional convolutional neural networks, vision transformers use self-attention to capture both local lesion characteristics and longer-range spatial relationships across an image. Our results demonstrate the potential of transformer-based architectures for discriminating melanoma from benign skin lesions and provide insights into their advantages and limitations.
Beyond benchmark classification, we will discuss the challenges involved in translating AI algorithms into practical skin cancer screening systems, including image resolution, lesion size, dataset bias, model generalizability, explainability, and clinical workflow integration. These results illustrate both the rapid progress of AI-based dermatologic image analysis and the opportunities for developing technologies capable of detecting skin cancers at early stages.
Speaker:
Xinming Huang
Professor, ECE Department, WPI
Bio:
Xinming Huang is a Professor of Electrical and Computer Engineering at Worcester Polytechnic Institute (WPI) and Director of the Embedded Computing Laboratory. His research interests include machine learning, computer vision, embedded design, and edge computing. His recent work focuses on AI algorithms for applications, including automated driving, wireless communications, and digital health.
Host: John McNeill