Analysis of class agnostic explanations for the prediction of Diabetic Retinopathy

Optical Coherence Tomography Angiography (OCTA) is a promising new imaging technique in the field of ophthalmology for diagnosing Diabetic Retinopathy (DR) in diabetic patients. Early detection and treatment of DR is crucial to preventing vision loss and blindness. With rising numbers of diabetes ca...

Descripción completa

Detalles Bibliográficos
Autor: Azimov, Shawn
Tipo de recurso: tesis de maestría
Fecha de publicación:2024
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/412464
Acceso en línea:https://hdl.handle.net/2117/412464
Access Level:acceso abierto
Palabra clave:Neural networks (Computer science)
Diabetic retinopathy
CNN
XAI
OCTA
DR
Optical Coherence Tomography Angiography
explainable artificial intelligence
convolutional neural networks
diabetic retinopathy
signature activation
saliency map
Xarxes neuronals (Informàtica)
Retinopatia diabètica
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial
Descripción
Sumario:Optical Coherence Tomography Angiography (OCTA) is a promising new imaging technique in the field of ophthalmology for diagnosing Diabetic Retinopathy (DR) in diabetic patients. Early detection and treatment of DR is crucial to preventing vision loss and blindness. With rising numbers of diabetes cases around the world, it is more important than ever to correctly diagnose and categorize DR. Novel approaches to image classification, such as convolutional neural networks (CNN) have proven to be very effective with many types of images, however, their adoption in the medical field hinges on the ability of the medical professionals to understand how these networks make their decisions. This thesis focuses on exploring explainable artificial intelligence (XAI) techniques to better understand how a CNN can classify real OCTA images of diabetes patients with varying levels of DR. In particular, the primary technique explored is Signature Activation, which can generate holistic and class-agnostic explanations of the CNN's decisions. Over the course of this thesis several CNN models were obtained utilizing different architectures and training hyperparameters. In addition, several explainability visualization techniques were developed and applied to these models at various layers in order to study how information propagates through the layers of deep neural networks and analyze how they arrive at their final outputs.