AI16 An explainable dataset and reporting framework for structured clinical reasoning in skin lesion assessment

Abstract

Assessment of skin lesions relies on structured reasoning frameworks such as ABCDE (asymmetry, border irregularity, colour variation, diameter, evolution). In routine practice, clinical reasoning, particularly when evaluating benign lesions that appear clinically suspicious, is often incompletely captured in referral documentation, contributing to variability in triage decisions and increased clinician workload. The aim of this study was to develop and evaluate a clinician-facing digital system that supports image annotation and speech-based input, using natural language processing (NLP) to structure clinical observations into ABCDE-based reasoning and generate automated dermatology reports. A software platform was developed to enable clinicians to annotate lesion images during assessment and to record spoken clinical observations. Speech-to-text transcription and NLP pipelines were used to categorize clinician input into structured ABCDE components. The system generated clinician-reviewable reports aligned with dermatology referral standards. Evaluation focused on the accuracy of ABCDE categorization, preservation of clinical intent and usability within routine workflows. The NLP system consistently mapped clinician annotations and spoken input into structured ABCDE reasoning, including features commonly cited when assessing clinically suspicious benign lesions. Automated reports were judged by clinicians to be clear, interpretable and reflective of clinician reasoning, while reducing documentation burden. This clinician-in-the-loop annotation and NLP system support structured capture of clinical reasoning during skin lesion assessment. By improving documentation quality and explainability, the platform enables safer AI-assisted workflows and more consistent dermatology referrals.