A deep learning model for diagnosis of inherited retinal diseases

Abstract

To evaluate the performance of a multi-input deep learning (DL) model in detecting two common inherited retinal diseases (IRDs), i.e. retinitis pigmentosa (RP) and Stargardt disease (STGD), and differentiating them from healthy eyes. This cross-sectional study includes 391 cases, consisting of 158 subjects with RP, 62 patients with STGD, and 171 healthy individuals. The image dataset is publicly available at http://en.riovs.sbmu.ac.ir/Access-to-Inherited-Retinal-Diseases-Image-Bank . Separate networks using the same hyperparameters were trained and tested on the dataset. Two single-input MobileNetV2 networks were employed for color fundus photography (CFP) and infrared (IR) images, and a multi-input MobileNetV2 network was applied using both imaging modalities simultaneously. The single-input MobileNetV2 achieved 94.44% diagnostic accuracy using CFP, and 94.44% accuracy employing IR images, respectively. The multi-input MobileNetV2 network outperformed both single-input networks with an accuracy of 96.3%. The impact of single-input and multi-input architectures was further evaluated on state-of-the-art neural network models and machine learning algorithms. The deep learning networks utilized in this study achieved high performance for detection of IRDs. Application of a multi-input network employing both CFP and IR image inputs improves the overall performance of the model and its diagnostic accuracy.

Keywords: Artificial intelligence; Deep learning architectures; Inherited retinal diseases; Machine learning algorithms; Retinitis pigmentosa; Stargardt disease.