Deep learning-based non-invasive differential diagnosis of eyelid basal cell and sebaceous gland carcinomas using...

Pathological examination, the current gold standard for differentiating eyelid basal cell carcinoma (BCC) and sebaceous gland carcinoma (SGC), is invasive, time-consuming, and often inaccessible in primary care hospitals. Therefore, a non-invasive, early differential diagnostic method for eyelid BCC and SGC is needed to reduce diagnostic delays and improve accuracy, which holds significant clinical value, especially for Asian populations.

Using 199 eyelid BCC and 171 eyelid SGC photographic images from Shanghai Ninth People's Hospital, Shanghai Jiao Tong University School of Medicine (2016-2022), we developed a ResNet50-based deep learning model. The model's performance was evaluated using classification accuracy and the F1 score and was compared with that of four ophthalmologists (three junior and one senior). To assess its clinical utility, we further evaluated the diagnostic accuracy of ophthalmologists with and without the model's assistance.

The developed model achieved a differential diagnosis accuracy of 0.892 (95% CI 0.821-0.937), with a sensitivity of 0.863 (95% CI 0.743-0.932) and a specificity of 0.917 (95% CI 0.819-0.964), outperforming all four ophthalmologists (junior: 0.703 (95% CI 0.612-0.780), 0.757 (95% CI 0.669-0.827), 0.811 (95% CI 0.729-0.873); senior: 0.874 (95% CI 0.799-0.923)). With the model's assistance, diagnostic accuracy improved by 14.1%, 16.6%, and 2.2% for junior ophthalmologists and by 1.0% for the senior ophthalmologist.

The developed model accurately differentiates eyelid BCC and SGC and effectively improves diagnostic performance, particularly for junior ophthalmologists. It may facilitate timely and appropriate treatment planning in clinical settings for eyelid BCC and SGC patients.