| Our study | Adamopoulos et al. [48] | Li et al. [49] |
---|---|---|---|
What was solved | • DL system can identify benign and malignant tumors through common clinical images, with a better performance than most ophthalmologists • Enable patients’ self-monitoring and assist in doctors’ clinical decision making when combining with smartphone App | • Provide full support that DL ANN can count as a powerful pattern recognition and classification tool and can be widely applied on tasks related to medical diagnosis on BCC | • The potential for promoting early detection and treatment of malignant eyelid tumors at screening stage • A roust performance in identifying malignant eyelid tumors from benign ones |
Methods and details | CNN for distinguishing malignant tumors Class weight and different epochs of 60 and 80 were applied to train the models without underfitting | ANN and CNN for distinguishing BCC | Faster-RCNN (Faster Region based CNN) for localizing tumors and CNN for distinguishing malignant tumors Each algorithm was trained for 80 epochs |
Outcomes (95% CI) | |||
 Accuracy | 0.889 (0.747–0.956) | Approximately 0.500–1.000 | 0.818 (0.773–0.862) |
 Sensitivity | 0.933 (0.702–0.988) | – | 0.915 (0.844–0.986) |
 Specificity | 0.857 (0.654–0.950) | – | 0.792 (0.739–0.845) |
 AUC | 0.966 (0.850–0.993) | – | 0.899 (0.854–0.934) |
Dataset | 339 photographic images from 255 patients, collected from Beijing Tongren Hospital, Beijing, China | 143 photographic images collected from the Clinic of Ophthalmology of the University Hospital at Heraklion, Crete, Greece | 1417 photographic images from 851 patients, collected from Ningbo Eye Hospital, Jiangdong Eye Hospital, and Zunyi First People’s Hospital |
Limitations | Cannot provide a specific subtype diagnosis based on images | May not be employed for other malignant eyelid tumors other than BCC | Cannot provide a specific subtype diagnosis based on images |