Intelligent Pathological Diagnosis of Melanocytic Lesions Using Deep Learning

Abstract

Objective: To develop a clinically applicable intelligent pathological diagnosis model for melanocytic lesions using deep learning (DL). Methods: Pathological slides from 218 malignant, 119 atypical, and 374 benign melanocytic lesions patients diagnosed between 2001 and 2018 were collected from Pathology Departments of Shanghai Ninth People's Hospital (Center 1), North Campus of Shanghai Ninth People's Hospital (Center 2), and Baoshan Branch of Shanghai First People’s Hospital (Center 3). These slides were digitized using Hamamatsu NanoZoomer S60 scanner, generating 981 whole-slide images (WSI). Among them, 745 images from Center 1 were designated as training set, 182 images as internal testing set, and 54 images from Centers 2 and Center 3 as external testing set. A DL-based stain style transfer strategy was designed to normalize stain variations in pathological images. Subsequently, by cascading DL prediction module with prediction results aggregation module, an intelligent diagnosis model was constructed to differentiate between melanocytic lesion types (malignant, atypical, and benign). The diagnosis model performance was validated on both internal and external testing sets. Results: The proposed DL-based stain normalization method obtained an average increasement of 0.072 and 11.20dB in Structural Similarity (SSIM) and Peak Signal-to-Noise Ratio (PSNR), respectively, compared to color-based and stain-based methods. The diagnosis model built upon this method achieved an accuracy of 94.12% on the internal testing set and over 90% on the external testing set. Conclusion: A highly generalized intelligent pathological diagnosis model for melanocytic lesions has been established, which holds significant implications for advancing clinical application of artificial intelligence (AI)-assisted pathological diagnosis.