Adaptive Segmentation of Oral Cavity Lesions via Color-based Custom Index and Differential Evolution

This work proposes an adaptive method for segmenting oral cavity lesions, designed to support AI-assisted diagnosis and improve early detection. The method is based on three basic components: (i) a segmentation index, inspired by the well-known Soil-Adjusted Vegetation Index (SAVI). In this work, the index is reinterpreted as a lesion index, exploiting subtle color variations between healthy and pathological tissues to identify clinically relevant areas of a lesion; (ii) a superpixel algorithm, which groups neighboring pixels into color-homogeneous regions to highlight its clinical morphology, and (iii) an evolutionary optimization algorithm, which tunes the parameters of both the index and the segmentation process to the specific clinical context. By leveraging intrinsic color features as diagnostic indicators, the approach enables robust segmentation even in the absence of annotated data. Preliminary experiments conducted on a representative set of oral lesion images demonstrate the effectiveness of the method. Segmentation performance was quantitatively evaluated using metrics such as Kappa coefficient, precision, showing promising results for clinical applications.