A Bradley-Terry-Luce model-based analysis of human perception and computer vision in identifying melanoma lesions
Melanoma is a deadly skin cancer, and early detection is critical for improving survival rates. Dermatologists typically rely on a visual scan to diagnose melanoma by assessing the primary perceptual characteristics of a skin lesion. The common ABCDE heuristic, for example, suggests observers check a lesion for shape (A)symmetry, (B)order irregularity, number of unique (C)olours, and (E)volution over time. Whilst this heuristic provides a practical guide, it is a limited approach. Firstly, all lesions vary and often contain only a subset of these features. Secondly, a combination of abnormal features can lead to a diagnosis, making the diagnostic process complicated and error-prone. Advanced computer vision algorithms (CVA) have emerged as a powerful approach to melanoma identification. CVAs can evaluate lesion features to generate highly accurate and objective assessments. However, despite CVA advancements, they can only be used in conjunction with an expert assessment. Thus, the perceptual expertise of dermatologists remains a critical component in the accurate and timely detection of melanoma. Our project aims to improve the early detection of melanoma by investigating the perceptual judgments of skin lesion colour and shape made by humans and comparing them with the feature representations generated by computer vision algorithms. We recruited non-expert participants online to complete a two-alternative forced-choice task using skin lesion images from the ISIC archive. Participants were instructed to choose the image that exhibited a greater frequency of unique colours in one condition and greater border regularity in another among the two images presented in a trial. We analysed the data using the Bradley-Terry-Luce (BTL) model to estimate each lesion image's relative "strengths" along these perceptual dimensions. We then compared these estimates to computer vision assessments of the same perceptual features. We discuss the methodological approach, preliminary results, and future directions.