Accurate classification of benign and malignant dermoscopy skin lesions using three deep learning models

Abstract

INTRODUCTION: Dermoscopy is a non-invasive technique used to evaluate suspicious skin lesions with a high magnification lens that allows users to assess structures in the reticular dermis. Our goal is to utilize dermoscopic images to train an artificial intelligence (AI) algorithm to recognize patterns and diagnose skin lesions as benign or malignant. Herein, we discuss our data collection strategy and results that were used to train the algorithm. METHODS: Chart review of patients at the University of Missouri occurred from 2021-2023. A total of 695 distinct dermoscopic images were collected, each tied to a diagnosis confirmed via biopsy. Information on age, gender, race, and the county was also recorded. RESULTS: Data set consisting of 357 patients demonstrated that the average age was 67. Patients consisted of 52.9% males and 47.1% females. When evaluating patients by race, 98% were White, 1% were Black/African American, and 1% identified as "Other race". Of the 14 counties represented, most patients resided in Boone (46%), Camden (4%), and Jefferson(4%). Of the 33 diagnoses evaluated, the top three were basal cell carcinoma (12%), malignant melanoma (10%), and dysplastic nevus (9%). When all the images were compared to biopsy results, 61% of diagnoses were benign and 39% were malignant. CONCLUSION: Recognition of subtle dermoscopic patterns takes years of training, which limits its use outside of dermatology. By using this data to train an AI algorithm, we hope to increase ease of access to this technology to help rural physicians and primary care providers triage skin lesions and spare patients from unnecessary biopsies.