A GRT examination of the ABCs for melanoma identification with comparison to a deep convolutional neural network.

Survival rates of malignant melanoma improve with early detection, but accurately distinguishing melanoma from an otherwise benign skin lesion is a difficult perceptual task. Rules-based heuristics, such as the ABCD criteria, direct novice observers to judge skin lesions for signs of melanoma according to the distinct perceptual features of shape asymmetry, border irregularity, colour variance, and diameter. Violations of a rule or multiple rules signal the observer to seek expert evaluation and offer a convenient albeit simplified approach to making such complex perceptual judgments. While helpful, the ABCD heuristic contains two primary limitations. First, it assumes an individual's capacity to judge these features independently. Second, research shows expert observers (i.e., dermatologists) judge skin lesions holistically rather than via distinct ABCD features. Deep Convolutional Neural Networks (DCNNs) are powerful machine learning techniques for image processing that are frequently applied to melanoma identification problems and can help reduce observer error. While powerful, these algorithms suffer from their black-box nature, ultimately limiting their utility in human-AI collaborative decision-making contexts. However, the final activation layers of a DCNN provide insight into the features necessary for the final classifier, which can be leveraged to inform human observers. Our project has two aims: (i) to examine if and how novice observers combine the perceptual information dimensions of shape asymmetry (A), border irregularity (B), and colour variance (C) and (ii) to explore the representational space of a DCNN trained for melanoma classification. We collected data across three experiments, each presenting a unique pairwise combination of features to be judged (AB, AC, and BC). We then modelled novice observers' perceptual space using GRT machinery. We found that participants failed to separate perceptual information across all dimensions, indicating that novice observers may rely on an overall 'ugliness' concept to drive their judgments. We then trained a DCNN for melanoma identification and identified the resultant features essential to this classifier. Our methodological approach and preliminary results are presented and discussed.