Purchase this article with an account.
Alexander Urich, Keyur Ranipa, Krunalkumar Ramanbhai Patel, Mary K Durbin, Christian Wojek, Alexander Freytag; Disease prediction from thickness maps derived from OCT scans using Machine Learning and only a handful of data. Invest. Ophthalmol. Vis. Sci. 2018;59(9):1731.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Surgeries related to the implantation of intraocular lenses (IOL) are wide-spread all around the world. A pre-requisite for a successful outcome is a healthy retina, which should be verified before surgery. Hence, an automated reliable health status assessment of the retina is highly desirable. In this paper, we aim at learning to predict the health status from thickness maps derived from a 3D OCT retina scan and compare different machine learning approaches to solve this task.
We conduct experiments on 3D OCT data obtained with a Cirrus HD-OCTTM 4000 (ZEISS, Dublin, CA) with 192 scans of normal patients and 213 scans of patients with AMD (non-normal). In addition to image data, the patient’s age is taken into account. Thickness maps, corresponding to an area of 6mm x 6mm (512x128 pixels), are generated by segmenting the ILM and RPE layers in each B-scan (Fig. 1). The data is split using 5-fold constrained cross-validation.For each split, we fine-tune a pre-trained CNN (InceptionV1) to the training set using TensorFlow (DeepCNN). For comparison, we train a shallow linear SVM on HOG features extracted from an 8x8 grid from the thickness maps (HOG-SVM). To investigate the potential bias from age, we append the normalized age to the HOG features (HOG-Age-SVM). All models are trained/evaluated on the same splits and hold-out test data. The final baseline is a cross-check of thickness map and age against a hold-out normative database, the de-facto standard procedure today (Normative). Since data from all splits are mutually exclusive, results from all splits are pooled together for the final comparison of all methods. We report ROC curves showing rates of true vs. false positives for varying thresholds. The area under the curve serves as an accuracy metric.
The results from Deep Learning (DeepCNN) and standard computer vision algorithms (HOG-SVM) with close to 95% AUC outperform the normative database (Normative) with 79% AUC (Fig. 2).
We evaluated normality classification from thickness maps computed from 3D OCT volumes. The comparison to an established normative database approach with shallow and deep learning algorithms clearly underlines the effectiveness of learning normality directly from annotated data.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.
Example thickness maps of normal (a) and diseased patients (b).
Results of predicting normality from thickness map derived from 3D OCT data.
This PDF is available to Subscribers Only