Enhancing dermatological diagnosis for differentiating actinic from seborrheic keratosis using deep learning model

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Differentiating Actinic keratosis (AK) from Seborrheic keratosis (SK) presents a diagnostic challenge due to visual similarities.
• Accurate classification is crucial for appropriate treatment and patient outcomes.

Purpose of the Study:

• To evaluate the efficacy of a deep learning (DL) model in assisting dermatologists to accurately classify AK versus SK lesions.
• To assess the impact of the DL model on diagnostic performance and decision-making.

Main Methods:

• A contrastive language-image pre-training (CLIP) model with ViT-B/16 architecture was trained on 2,307 patient cases.
• The model was validated across three independent datasets. Dermatologists' classifications were compared before and after utilizing the DL model's predictions.
• Diagnostic performance was measured using Area Under the Receiver Operating Characteristic Curve (AUC), Net Reclassification Index (NRI), and Total Integrated Discrimination Index (IDI).

Main Results:

• The DL model achieved AUCs ranging from 0.85 to 0.89 across training and validation cohorts.
• Dermatologist 1's diagnostic performance improved from 0.77 to 0.80 (AUC), with significant NRI (0.10) and IDI (0.14) changes.
• Dermatologist 2 showed a substantial improvement from 0.69 to 0.79 (AUC), with significant NRI (0.19) and IDI (0.27) changes.

Conclusions:

• The DL model significantly enhances dermatologists' accuracy in differentiating AK from SK.
• The AI tool has the potential to reduce diagnostic subjectivity and aid in the early detection of precancerous lesions.
• Implementation of DL models could transform dermatological diagnostic and therapeutic practices.