OCT Scan - Image Classification

Project description

Retinal optical coherence tomography (OCT) is an imaging technique used to capture high-resolution cross sections of the retinas of living patients. It has become a standard imaging modality for diagnosis and treatment management of disease leading to vision loss with nearly 30 million scans acquired annually (Swanson and Fujimoto, 2017). Age related diseases such as age-related macular degeneration (AMD) and diabetic macular edema (DME), choroidal neovascularization (CNV) are the leading causes of blindness worldwide (Varma et al., 2014, Wong et al., 2014), and are likely to become more prevalent with an aging population. Early diagnosis of CNV and DME are crucial, as delayed treatment could lead to irreversible loss of vision.

We constructed an ensemble CNN classifier to diagnose the most common retinal diseases and improve upon the results presented by Kermany et. al. 2018. Our ensemble method consists of four CNNs widely used in image classification tasks (VGG16, ResNet50, Xception, InceptionV3). These networks were pretrained on the ImageNet dataset, then fine tuned on 1000 OCT images (250 NORMAL, 250 CNV, 250 DME, 250 DRUSEN) labeled by a team of expert clinicians. By calculating the mean class probability across each classifier, we were able to achieve superior results to any individual classifier as well as the previous published results.

On an independent test set of 1000 OCT images (250 NORMAL, 250 CNV, 250 DME, 250 DRUSEN), our classifier achieved a 4 class accuracy of 96.4%, which is a 3% improvement on what was reported by Kermany et. al. 2018 on the same test data. To reflect the potential clinical use of these predictions, labels were also binarized into two groups: URGET (CNV, DME) and NON-URGENT (NORMAL, DRUSEN) where the ensemble CNN achieved an accuracy of 97.5%, sensitivity of 98.6%, specificity of 96.5%, and ROC AUC of 99.8%.

The data used for training and testing may be found here and here.