Abstract
Purpose :
Recent studies have shown the relevance of deep learning for detecting frequent pathologies, such as diabetic retinopathy (DR), age-related macular degeneration (AMD) and glaucoma, in fundus photographs. However, due to insufficient training data, rare pathologies cannot be detected well through deep learning. We believe that ignoring rare pathologies is a strong limitation of current artificial intelligence (AI) systems. Therefore, this study investigates the use of few-shot learning, a far less data-demanding paradigm, to also detect rare pathologies: the idea is to reuse deep learning models previously trained to detect frequent pathologies.
Methods :
The AI was developed and evaluated using deidentified DR screening examinations from the OPHDIAT screening network in Paris, France. In this network, each screening examination is analyzed by one of seven certified ophthalmologists. Both structured and free-form examination reports were analyzed retrospectively: a list of 41 pathologies and pathological signs was established and ground truth was obtained for these 41 diagnoses on a subset of 115,159 images. The AI was developed using 32% of this dataset. Performance was evaluated in the remaining 68%, using the area under the ROC curve (Az).
Results :
Out of the 41 investigated pathologies and pathological signs, 37 could be detected well (Az>0.8). This includes degenerative myopia (Az=0.9999), central retinal vein occlusion (Az=0.9873), chorioretinitis (Az=0.9913), retinis pigmentosa (Az=0.9984), papilledema (Az=0.9510) and anterior ischemic optic neuropathy (Az=0.9534). Of course, this also includes the most frequent pathologies: DR (Az=0.9882), glaucoma (Az=0.9733), cataract (Az=0.9834) and AMD (Az=0.9916). Signs that were poorly detected include shunts (Az=0.7586) and emboli (Az=0.7946). Additionally, the detected anomalies can be highlighted in images.
Conclusions :
We have designed an AI able to detect 37 pathologies and pathological signs in fundus photographs. This AI, which was developed and evaluated on a DR screening dataset, is expected to improve the usefulness and reliability of automated DR screening in the diabetic population: even if DR is not detected, a warning and a detailed report can be generated should another problem be detected. This AI might also be applied to mass screening of eye pathologies in the general population: future studies will thus investigate its performance in that more general context.
This is a 2020 ARVO Annual Meeting abstract.