Automatic glaucoma screening with low-cost devices

• Autor(a) principal: Neto, Alexandre Henrique da Costa
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10348/10401
• topic: Machine learning; glaucoma screening

Glaucoma is a silent disease that does not show symptoms until it is too late, causing irreversible blindness. Broader screening programs are limited by specialised teams and equipment's high costs. Current machine learning methods can help glaucoma screening, lower the cost, and extend it to larger populations. In less developed countries, medical centres capable of screening glaucoma are scarce. Low-cost lenses attached to mobile devices can increase the screening ampleness alerting patients to go to medical centres earlier for a more profound evaluation. These devices can capture images with enough quality and run machine learning models (e.g. convolutional neural networks, CNNs) reported by the scientific community to achieve excellent results in medical imaging. However, these solutions are not yet in use and need further study. In this work, we explore and compare the contributions of state-of-the-art classification and segmentation CNN's methods for automatic glaucoma screening with retinal images acquired both by retinographers and low-cost lenses attached to mobile devices. We used classification methods with the Xcpetion, ResNet152 V2 and the Inception ResNet V2 models. To support glaucoma classifiers predictions, we produced and analysed the model’s activation maps, allowing specialists to understand and trust the results achieved. In the clinical, the glaucoma assessment is commonly based on the Cup to Disc Ratio (CDR) criterium, an evident indicator for specialists. For this reason, additionally, we use the U-Net architecture with the Inception ResNet V2 and Inception V3 models as the backbone to segment and estimate CDR. For both tasks, the models were trained and evaluated with high-quality retinal images from public databases (RIM-ONE, DRISHTI-GS and REFUGE). The classification models were trained and evaluated additionally with a private dataset, with retinal low-quality images acquired with a smartphone with a low-cost lens coupled. The classification models achieved performances with similar results to the state-of-theart and producing activation maps gives interpretability to analysed and discuss the output, making it easier for the ophthalmologist to understand the model’s decision. The optic disc (OD) and cup segmentation on high-quality images (public datasets) reached high performances and the glaucoma classification through CDRs reached state-of-theart results proving the significance of the CDRs indicators to detect glaucoma. The representations of the OD and cup outlines help clinicians to an easier examination based on CDRs criterium.The same methods of classification were applied in the low-quality private database with slightly lower results but showing a contribution to the use of cheaper lens attached to mobile devices. This type of lens can expand and turn mass screening more accessible in remote areas.