Author(s) :

Shreya Talukder.

Volume/Issue :

Abstract :

Skin cancer is the most common form of cancer worldwide, and early detection plays a critical role in improving survival outcomes. While early-stage melanoma has a five-year survival rate of approximately 99%, this rate declines significantly at later stages, emphasizing the need for timely and accessible diagnostic tools. However, access to dermatological care remains limited for billions of people worldwide due to cost, geography, and time constraints. In this work, we present AutoLesion, an affordable and accessible artificial intelligence–based system for the preliminary assessment of skin lesion malignancy using a novel multimodal approach. Unlike prior methods that rely solely on dermoscopic or highresolution imagery, AutoLesion integrates cell phone images with clinical and symptomatic metadata through a fine-tuned vision–language model (VLM). This joint utilization of visual and clinical information captures indications of malignant skin lesions that are often overlooked by image-only models. We further introduce a test-time compute strategy to improve prediction accuracy and reliability. Experimental evaluation on the ISIC (International Skin Imaging Collaboration) skin lesion archive demonstrates that the proposed approach outperforms dermoscopic image-only baselines, supporting the effectiveness of multimodal diagnosis from consumer-grade imagery. Even though these systems can improve early detection, especially in disadvantaged regions, these advantages are outweighed by practical and ethical concerns such as algorithmic bias, data privacy, and the need for human oversight. These considerations underscore the importance of responsible development and deployment of AIassisted medical diagnostic tools.