In Retina (Philadelphia, Pa.)
PROPOSE : A screening protocol for cytomegalovirus retinitis (CMVR) by fundus photography was generated and the diagnostic accuracy of machine learning technology for CMVR screening in HIV patients was investigated.
METHODS : One hundred and sixty-five eyes of ninety HIV positive patients were enrolled and evaluated for CMVR with binocular indirect ophthalmoscopy. Then, a single central field of the fundus image was recorded from each eye. All images were then interpreted by both machine learning models, generated by using the Keras application, and by a third-year ophthalmology resident. Diagnostic performance of CMVR screening using a machine learning model and the third-year ophthalmology resident were analyzed and compared.
RESULTS : Machine learning model, Keras application (VGG16) provided 68.8% (95%CI=50-83.9%) sensitivity and 100% (95%CI= 97.2-100%) specificity. The program provided accuracy of 93.94%. While the sensitivity and specificity for the third-year ophthalmology grading were 67.7% (95%CI=48.6-83.3%) and 98.4% (95%CI=94.5-99.8%). The accuracy for CMVR classification was 89.70%. When considering for sight threatening retinitis in zone 1 and excluded zone 2 and 3, the machine learning model provided high sensitivity of 88.2% (95%CI=63.6-98.5%) and high specificity of 100% (95%CI= 97.2-100%).
CONCLUSIONS : This study demonstrated the benefit of the machine learning model VGG16 which provided high sensitivity and specificity for detecting sight threatening CMVR in HIV positive patients. This model is a useful tool for ophthalmologists in clinical practice for preventing blindness from CMVR, especially during the COVID-19 pandemic.
Srisuriyajan Pitchapa, Cheewaruangroj Nontawat, Polpinit Pattarawit, Laovirojjanakul Wipada