Diagnosis of Skin Disease in Moderately to Highly Pigmented Skin by
Artificial Intelligence
Abstract
Background: Triage of patients with skin diseases often
includes an initial assessment by a nurse or general practitioner,
followed by a dermatologist. Artificial intelligence (AI) systems have
been reported to improve clinician ability to diagnose and triage skin
conditions. Previous studies have also shown that diagnosis in patients
with skin of color can be more challenging.
Purpose: This study seeks to determine the performance of AI in
the screening and triage of benign-neoplastic, malignant-neoplastic, and
non-neoplastic skin conditions for Fitzpatrick IV-VI skin types.
Methods: A set of 163 non-standardized clinical photographs of
skin disease manifestations from patients with Fitzpatrick IV-VI skin
types were obtained through a publicly available dataset (Scale AI and
MIT Research Lab, “Fitzpatrick 17 Dataset”). All photos were diagnosed
by a specialist and categorized into three disease classes:
benign-neoplastic, malignant-neoplastic, or non-neoplastic. There were
23, 14, and 122 cases of each disease class, respectively.
Results: Overall, the AI was able to classify the disease
classes with a high degree of accuracy for the Top 1 diagnosis
(86.50%). Based on its first prediction, the AI demonstrated the
greatest accuracy when classifying non-neoplastic conditions (90.98%),
high accuracy of detecting malignant-neoplastic conditions (77.78%),
and moderate accuracy of classifying benign-neoplastic conditions
(69.57%).
Conclusion: The AI had an overall accuracy of 86.50% in
diagnosing skin disease in Fitzpatrick IV-VI skin types. This is an
improvement over reported clinician diagnostic accuracy of 44.3% in
darker skin types. Incorporating AI into front-line screening of skin
conditions could thereby assist in patient triage and shorten the time
to accurate diagnosis.