Morphometric Image Analysis comparison of basal cell carcinomas & melanomas of the head & neck versus other sites

Background Dermoscopic appearance of basal cell carcinomas (BCC) and melanomas guide diagnosis and may be influenced by anatomic location. Objective We aimed to investigate the morphometric differences between anatomic sites of the head and neck(H&N) versus other body sites in basal cell carcinomas and melanomas and how these differences may impact neural network model accuracy. Methods Morphometric image analysis of the BCCs(n=422) and melanomas(n=868) from the publicly available HAM10000 dataset was performed using an open-source image analysis software. Univariate and multivariate statistical analysis was done to identify differences between H&N and other anatomic sites. The multifactorial data was further interrogated with dimensionality reduction techniques. Randomly generated neural networks with different anatomic proportions of training data were used to assess how these features influence diagnostic accuracy. Results Fifty-three univariate and 11 multivariate features were found to be statistically significant in the BCC group(P<0.05). Thirteen univariate and 8 multivariate features were statistically significant in the melanoma group(P<0.05). Dimensionality reduction via linear discriminant analysis of the BCC groups revealed modest separation of the data by anatomical site. Melanomas appeared to be more homogenous across anatomic sites. Morphometric features significantly influenced neural network accuracies in classifying BCCs, with a mix of H&N and other site training data resulting in the best models(P<0.05). Conclusion Morphometric features and the anatomic composition of BCC training data may influence the accuracy of AI models. Model developers should be aware of the anatomic distribution of their training datasets and report site-specific validation metrics.