Precise higher-order reflectivity and morphology models for early diagnosis of diabetic retinopathy using OCT images

Abstract

This study proposes a novel computer assisted diagnostic (CAD) system for early diagnosis of diabetic retinopathy (DR) using optical coherence tomography (OCT) B-scans. The CAD system is based on fusing novel OCT markers that describe both the morphology/anatomy and the reflectivity of retinal layers to improve DR diagnosis. This system separates retinal layers automatically using a segmentation approach based on an adaptive appearance and their prior shape information. High-order morphological and novel reflectivity markers are extracted from individual segmented layers. Namely, the morphological markers are layer thickness and tortuosity while the reflectivity markers are the 1st-order reflectivity of the layer in addition to local and global high-order reflectivity based on Markov-Gibbs random field (MGRF) and gray-level co-occurrence matrix (GLCM), respectively. The extracted image-derived markers are represented using cumulative distribution function (CDF) descriptors. The constructed CDFs are then described using their statistical measures, i.e., the 10th through 90th percentiles with a 10% increment. For individual layer classification, each extracted descriptor of a given layer is fed to a support vector machine (SVM) classifier with a linear kernel. The results of the four classifiers are then fused using a backpropagation neural network (BNN) to diagnose each retinal layer. For global subject diagnosis, classification outputs (probabilities) of the twelve layers are fused using another BNN to make the final diagnosis of the B-scan. This system is validated and tested on 130 patients, with two scans for both eyes (i.e. 260 OCT images), with a balanced number of normal and DR subjects using different validation metrics: 2-folds, 4-folds, 10-folds, and leave-one-subject-out (LOSO) cross-validation approaches. The performance of the proposed system was evaluated using sensitivity, specificity, F1-score, and accuracy metrics. The system’s performance after the fusion of these different markers showed better performance compared with individual markers and other machine learning fusion methods. Namely, it achieved $$96.15\%$$ 96.15 % , $$99.23\%$$ 99.23 % , $$97.66\%$$ 97.66 % , and $$97.69\%$$ 97.69 % , respectively, using the LOSO cross-validation technique. The reported results, based on the integration of morphology and reflectivity markers and by using state-of-the-art machine learning classifications, demonstrate the ability of the proposed system to diagnose the DR early.

Predicting CT-Based Coronary Artery Disease Using Vascular Biomarkers Derived from Fundus Photographs with a Graph Convolutional Neural Network

“…We decided to manually annotate the vasculature and focused on developing the pipeline for vascular feature extraction and building the classification models. In the future, when more fundus images are available, a fully automatic blood vessel segmentation and modeling technique will be required for extracting the vascular structure [ 40 , 41 , 42 ]. Finally, we only included similarity scores based on two types of demographic information (i.e., age and gender).…”

The Role of Medical Image Modalities and AI in the Early Detection, Diagnosis and Grading of Retinal Diseases: A Survey

“…For OCT images, different methods have been applied. For example, Sharafeldeen et al [ 145 ] detected DR from OCT images using features that were extracted from 12 retinal layers, including the thickness, tortuosity, and reflectivity of each layer. Two-level neural networks were used for classification.…”

The Role of 3D CT Imaging in the Accurate Diagnosis of Lung Function in Coronavirus Patients

“…In order to capture the inhomogeneity that may be caused by COVID-19 infection, a Markov-Gibbs Random Field (MGRF) model [31][32][33] is utilized, which is one of the mathematical models that shows a high ability to capture the inhomogeneity in the virtual appearance model. An instance of an MGRF is specified by an interaction graph, defining which voxels are considered neighbors, and a Gibbs Probability Distribution (GPD) on that graph, which gives the joint probability density of gray levels in a voxel neighborhood.…”