IMPORTANCE: Artificial intelligence (AI) algorithms are under development for use in diabetic retinopathy photo screening pathways. To be clinically acceptable, such systems must also be able to classify other fundus abnormalities and clinical features at the point of care. BACKGROUND: We aimed to develop an AI system that can detect several fundus pathologies and report relevant clinical features. DESIGN: Convolutional neural network training with retrospective data set. PARTICIPANTS: Colour fundus photos were obtained from publicly available fundus image databases. METHODS: Images were uploaded to a web-based AI platform for training and validation of AI classifiers. Separate classifiers were created for each fundus pathology and clinical feature. MAIN OUTCOME MEASURES: Accuracy, sensitivity, specificity and area under receiver operating characteristic curve (AUC) for each classifier. RESULTS: We obtained 4435 images from publicly available fundus image databases. AI classifiers were developed for each disease state above. Although statistical performance was limited by the small sample size, average accuracy was 89%, average sensitivity was 75%, average specificity was 89% and average AUC was 0.58. CONCLUSION AND RELEVANCE: This study is a proof-of-concept AI system that could be implemented within a diabetic photo-screening pathway. Performance was promising but not yet at the level that would be required for clinical application. We have shown that it is possible for clinicians to develop AI classifiers with no previous programming or AI knowledge, using standard laptop computers.
IMPORTANCE: Artificial intelligence (AI) algorithms are under development for use in diabetic retinopathy photo screening pathways. To be clinically acceptable, such systems must also be able to classify other fundus abnormalities and clinical features at the point of care. BACKGROUND: We aimed to develop an AI system that can detect several fundus pathologies and report relevant clinical features. DESIGN: Convolutional neural network training with retrospective data set. PARTICIPANTS: Colour fundus photos were obtained from publicly available fundus image databases. METHODS: Images were uploaded to a web-based AI platform for training and validation of AI classifiers. Separate classifiers were created for each fundus pathology and clinical feature. MAIN OUTCOME MEASURES: Accuracy, sensitivity, specificity and area under receiver operating characteristic curve (AUC) for each classifier. RESULTS: We obtained 4435 images from publicly available fundus image databases. AI classifiers were developed for each disease state above. Although statistical performance was limited by the small sample size, average accuracy was 89%, average sensitivity was 75%, average specificity was 89% and average AUC was 0.58. CONCLUSION AND RELEVANCE: This study is a proof-of-concept AI system that could be implemented within a diabetic photo-screening pathway. Performance was promising but not yet at the level that would be required for clinical application. We have shown that it is possible for clinicians to develop AI classifiers with no previous programming or AI knowledge, using standard laptop computers.
Authors: Michele Lanza; Robert Koprowski; Rosa Boccia; Katarzyna Krysik; Sandro Sbordone; Antonio Tartaglione; Adriano Ruggiero; Francesca Simonelli Journal: Front Med (Lausanne) Date: 2020-12-11