Resumen de Deep Learning-based Methods for Extracting Fundus Image Landmarks and Signs of Eye Diseases

Mohammed Yousel Salem Ali

• español

  La retinopatía diabética (RD) es la causa más típica de pérdida visual en adultos en edad laboral. En 2040, se prevé que más de 200 millones de personas tendrán RD. El glaucoma afecta a unos 75 millones de personas en todo el mundo y se le llama el ladrón silencioso de la vista. Por lo tanto, el diagnóstico precoz de glaucoma y RD requiere un procedimiento de detección eficaz. Para ello, los centros médicos realizan revisiones oftalmológicas periódicas a los pacientes, especialmente a los diabéticos, para minimizar el riesgo de ceguera. Examinar las imágenes del fondo de ojo requiere mucho trabajo, mucho tiempo, es costoso y es propenso a errores. Por lo tanto, es crucial desarrollar un sistema de diagnóstico asistido por computadora (CAD) que analice imágenes de fondo de ojo para ayudar a los oftalmólogos. Las técnicas de visión por computadora, como el aprendizaje profundo y especialmente las redes neuronales convolucionales (CNN), han mejorado significativamente el rendimiento de los sistemas CAD.

• català

  La retinopatia diabètica (RD) és la causa més típica de pèrdua visual en adults en edat de treballar. El 2040, es preveu que més de 200 milions de persones tindran DR. El glaucoma afecta uns 75 milions d'individus a tot el món i s'anomena el lladre silenciós de la vista. Per tant, el diagnòstic precoç del glaucoma i la RD requereix un procediment de cribratge eficient. Per tant, els centres mèdics realitzen regularment revisions oculars als pacients, especialment als diabètics, per minimitzar el risc de ceguesa. L'examen d'imatges del fons de l'ull requereix molta expertesa, temps, cost i genera errors. Per tant, és crucial desenvolupar un sistema de diagnòstic assistit per ordinador (CAD) que analitzi les imatges del fons ocular per ajudar els oftalmòlegs. Les tècniques de visió per ordinador, com ara l'aprenentatge profund i especialment les xarxes neuronals convolucionals (CNN), han millorat significativament el rendiment dels sistemes CAD. Aquesta tesi utilitza tècniques de visió per ordinador per considerar quatre tasques oculars: segmentació del disc òptic, detecció de glaucoma, segmentació de lesions degudes a DR com exudats o microaneurismes a partir d'imatges de fons d'ull.

• English

  The early detection and treatment of eye illnesses such as Diabetic retinopathy (DR) and Glaucoma disease in the retina are crucial to avoid vision loss. However, manual detection of small lesions in the fundus image is a painstaking process that consumes ophthalmologists' time and effort. The complex structure of lesions, various sizes, differences in brightness, and the inter-class similarity with other fundus tissues make it more challenging, even for ophthalmology experts. Therefore, it would be more beneficial to have a computer-aided diagnosis (CAD) system that can automatically outline the possible disease regions to the doctor.

  In this thesis, we consider four different tasks on the eye: optic disc segmentation, glaucoma detection, segmenting exudate lesions, and segmenting other kinds of retinal eye DR lesions from fundus images. These tasks are extremely challenging due to several sources of variability in the image-capturing processes, the complex structure, similarities of lesions, and a small amount of annotated images.

  Optic disc (OD) carries essential information linked to DR and glaucoma. Therefore, The first contribution in this thesis is a deep learning-based system for OD segmentation based on an ensemble of efficient semantic segmentation models. The aggregation was performed with the ordered weighted averaging operators on the different deep learning models. We propose an andness-directed set of weights to give a different contribution to the models according to their performance results. Tests were done with in-house dataset from Hospital Sant Joan de Reus.

  The second contribution is an efficient CAD system for diagnosing glaucoma utilizing transfer learning and fuzzy aggregation operators. Specifically, the proposed CAD system includes three stages: 1) Detection of the region of interest of the optic disc using an efficient deep learning network, 2) Classification based on different pre-trained deep convolutional neural networks and support vector machines, and 3) Use of fuzzy aggregation operators to fuse the predictions of glaucoma classifiers: ordered weighted average, weighted power mean, and exponential mean. Three public datasets were used: DRISHTI-GS1, RIM-ONE, and REFUGE.

  Next, the thesis deals with the automatic detection of DR lesions in the eye.

  The third contribution of the thesis is two novel exudate segmentation methods. The first method is based on a dual-decoder boosted network with a pre-trained ImageNet ResNet50 encoder, Atrous Spatial Pyramid Pooling (ASPP), and gated skip connections (GSCs). The second method is based on two multi-scale modules with ImageNet MobileNet encoder, ASPP, and GSCs. The effectiveness of the two methods was assessed on the IDRiD dataset. They obtained promising results outperforming numerous previous methods.

  As a fourth contribution, the thesis proposes an integrated retinal lesion segmentation model called LezioSeg, for segmenting the four typesof lesions: hard exudates (EX), soft exudates (SE), hemorrhages (HE), and microaneurysms (MA). It comprised four main elements: two multi-scale mules, ASPP at the neck of the network and multi-scale attention (SAT) unit after the decoder of the network, MobileNet backbone encoder, and several GSCs in the decoder block. It is 10 Million parameters, much lighter than those models that use ResNets or VGGNets backbones or those models that use the dual networks and leads to high segmentation performance of retinal eye lesions with the IDRiD and E-ophtha datasets. Extensive experiments had shown that the method showed superiority over other models and a competitive performance when it was generalized on the DDR dataset, which took in different imaging conditions. It can work well in the cases of medical images with small objects from different fields.