IS ARTIFICIAL INTELLIGENCE HELPFUL ON DERMOSCOPY EVALUATION OF ERYTHEMATOUS SKIN TUMOURS?

Abstract

Background: Erythematous skin tumours, benign and malignant, can be challenging both clinically and dermoscopically. Artificial Intelligence (AI) has shown encouraging results in dermatology, especially on the recognition of skin cancer. AI effectiveness and value in evaluating non or less pigmented lesions, namely rare tumor types, is still unclear and under investigation.The aim of this prospective study was to compare AI diagnostic accuracy with expert dermatologists and histopathological results. Methods: Clinical and dermoscopic images of erythematous lesions (n=111)were collected from 100 patients observed in the Dermatology Department of the Portuguese Oncology Institute of Coimbra (IPO), in Portugal, from January to August 2024. It was used an Handyscope attached to a mobile phone with AI (Fotofinder Systems®) to collect and analyse the images. Suspected malignant lesions were excised and a histopathological examination was disclosed. The uncertain lesions on dermoscopy or clinical examination were also excised or biopsied for diagnostic purposes.The undoubtful benign lesions identified by three expert dermatologists,after examining the clinical and dermoscopic images were not excised. All images were confronted with AI evaluation. Results: Concerning confirmed malignant lesions and AI, 96.6% of the dermoscopic images only were considered high risk. Based on clinical images, AI was not able to classify 49.2% of the malignant lesions and 47.5% were considered high risk for malignancy. Regarding benign lesions and AI, 52% of dermoscopic and 15.4% of macroscopic images were classified as high risk. AI could not give an evaluation on most of the clinical images (73.1%).The panel of dermatologists showed precision and agreement on their diagnosis.The diagnosis accuracy score of AI on dermoscopic images was 74%, the sensitivity was 96.6% and specificity was 48.1%. Discussion:AI was effective in identifying malignant erythematous lesions on dermoscopy images but showed limitations on the evaluation of clinical images. It was responsible for high rates of false-positive findings respecting benign lesions, both on dermoscopy and clinical images. In this study AI demonstrated some weaknesses suggesting that there is still necessary improvement and validation of algorithms in real-world clinical scenarios.

Datum der Bewilligung	2025
Originalsprache	Englisch
Gradverleihende Hochschule	Medizinische Universität Graz
Betreuer/-in	Andreas Blum (Betreuer*in)