AI11 Explainable multiclass artificial intelligence for safe identification and subtyping of benign skin lesions in suspected cancer pathways

Abstract

Urgent skin cancer referral pathways in the UK are characterized by very high volumes of benign disease. Many referrals arise from benign lesions that raise clinical concern for malignancy, reflecting diagnostic uncertainty in primary care and placing avoidable pressure on dermatology services. Our aim was to develop and evaluate an explainable artificial intelligence decision-support approach designed to safely identify benign lesions and provide benign subtype differentiation for lesions commonly referred under suspected cancer pathways (including seborrhoeic keratoses, intradermal naevi, cherry angiomas and other benign lesions). Publicly available, anonymized dermoscopic and close-up datasets were used. The system combines segmentation-guided image analysis with a two-stage classification framework: (i) a safety gate that separates suspicious cases from benign; and (ii) benign subtype differentiation for nonsuspicious cases. Explainability outputs were generated alongside predictions, including structured visual and text-based reasoning derived from lesion segmentation (shape/size and colour-region analysis) and supportive nondiagnostic dermoscopic structure cues. The model achieved high safety for benign identification (specificity ≥ 87%, negative predictive value ≥ 87%, area under the receiver operating characteristic curve ≥ 90% for nonsuspicious classifications). Explainability outputs were evaluated using structured clinician feedback, including clarity, usefulness for manage-versus-refer decisions, confidence calibration and interpretation time (five-point Likert scales). The outputs demonstrated consistent, clinician-­interpretable reasoning aligned with observable lesion characteristics. An explainable multiclass approach focused on benign identification and subtyping may reduce unnecessary urgent referrals while maintaining patient safety, with outputs that support transparent clinical reasoning in primary care workflows.