Diagnostic assessment of deep learning for melanocytic lesions using whole-slide pathological images
Problem
Discriminating melanoma from nevus on whole-slide pathological images is difficult and depends on expert dermatopathologists.
Why it matters
Misdiagnosis of melanoma carries life-or-death consequences and expert pathology time is limited.
What the AI did
A deep learning algorithm was trained on 781 whole-slide images to classify melanoma versus nevus, then tested on an independent hospital's images and compared against 7 expert dermatopathologists.
Results
Sensitivity 100% at specificity 94.7%, AUC 0.99 on independent test set of 104 WSIs; dermatopathologists achieved mean sensitivity 95.1% (confidence interval truncated).
Impact
Independent-hospital test set and expert comparison support the stated high diagnostic performance.
Record
| Venue | Translational Oncology |
|---|---|
| Published | 2021 |
| Work type | article |
| Authors | Wei Ba, Rui Wang, Guang Yin, Zhigang Song, Jinyi Zou, Cheng Zhong, Jingrun Yang, Guanzhen Yu et al. |
| Citations | 37 (OpenAlex, at import time) |
| References | 27 |
| Model or method | not stated in source |
| Organization | PLA General Hospital (training/validation data); Tianjin Chang Zheng Hospital (independent test set) |
| Primary source | https://doi.org/10.1016/j.tranon.2021.101161 |
| Open access | gold |
| Keywords | melanoma, whole-slide images, deep learning, dermatopathology, diagnostic accuracy |
Importance rating
significant — thousands to millions affected
Melanoma diagnosis is high-stakes pathology where errors cost lives and expert dermatopathologist time is scarce; abstract shows expert-level performance on independent hospital data but states no disease burden scale.