Cystoid Fluid Color Map Generation in Optical Coherence Tomography Images Using a Densely Connected Convolutional Neural Network

• Main Authors: Vidal, Placido L., de Moura, Joaquim, Novo, Jorge, Ortega, Marcos
• Format: Conference Proceeding
• Language: English
• Subjects: Biomedical imaging; Feature extraction; Image color analysis; Image segmentation; Information and communication technology; Pathology; Retina
• ISSN: 2161-4407
• DOI: 10.1109/IJCNN.2019.8852208

Optical Coherence Tomography (OCT) is a medical imaging modality that is currently the focus of many advancements in the field of ophthalmology. It is widely used to diagnose relevant diseases like Diabetic Macular Edema (DME) or Age-related Macular Degeneration (AMD), both among the principal causes of blindness. These diseases have in common the presence of pathological cystoid fluid accumulations inside the retinal layers that tear its tissues, hindering the correct vision of the patient. In the last years, several works proposed a variety of methodologies to obtain a precise segmentation of these fluid regions. However, many cystoid patterns present several difficulties that harden significantly the process. In particular, some of these cystoid bodies present diffuse limits, others are deformed by shadows, appear mixed with other tissues and other complex situations. To overcome these drawbacks, a regional analysis has been proven to be reliable in these problematic regions. In this work, we propose the use of the DenseNet architecture to perform this regional analysis instead of the classical machine learning approaches, and use it to represent the pathological identifications with an intuitive color map. We trained, validated and tested the DenseNet neural network with a dataset composed of 3247 samples labeled by an expert. They were extracted from 156 images taken with two of the principal OCT devices of the domain. Then, this network was used to generate the color map representations of the cystoid areas in the OCT images. Our proposal achieved robust results in these regions, with a satisfactory 97.48% ± 0.7611 mean test accuracy as well as a mean AUC of 0.9961 ± 0.0029.