Objectives. To evaluate CNN models’ performance of identifying the clinical images of basal cell carcinoma (BCC) and seborrheic keratosis (SK) and to compare their performance with that of dermatologists. Methods. We constructed a Chinese skin diseases dataset which includes 1456 BCC and 1843 SK clinical images and the corresponding medical history. We evaluated the performance using four mainstream CNN structures and transfer learning techniques. We explored the interpretability of the CNN model and compared its performance with that of 21 dermatologists. Results. The fine-tuned InceptionResNetV2 achieved the best performance, with an accuracy and area under the curve of 0.855 and 0.919, respectively. Further experimental results suggested that the CNN model was not only interpretable but also had a performance comparable to that of dermatologists. Conclusions. This study is the first on the assistant diagnosis of BCC and SK based on the proposed dataset. The promising results suggested that CNN model’s performance was comparable to that of expert dermatologists.

No abstract