Exploring mesh-based features for enhanced dermoscopic lesion classification

Abstract

Computational methods in digital image analysis have made significant progress in facilitating early diagnosis across various medical specialties, especially in dermoscopy. This study introduces a mesh-based classification approach that uses geometric and color information extracted from pigmented skin lesions to aid in the description and prediction of melanoma. The proposed method initially constructs Delaunay meshes to model segmented lesions under specific settings. Subsequently, mesh-derived and graph-based features are extracted and selected to feed machine learning algorithms. An experimental evaluation of our proposal on 274 public dermoscopies with melanoma and nevus abnormalities indicates that our approach achieved competitive performance against deep and shallow learning approaches. In particular, by using fewer training images than the literature methods, the best proposal configurations yielded an average accuracy of 86% and reduced the number of features down to 24. Notably, mesh configurations with a variable number of vertices around the lesion border and a fixed number within the lesion demonstrated promising results. The mesh-based classification method proposed in this work shows promise in dermoscopy and presents potential for application in other medical image classification domains.