Abstract
Purpose: :
To develop an automated method to identify the normal macula and four macular pathologies (macular hole (MH), macular edema (ME), epiretinal membrane (ERM), age-related macular degeneration (AMD)) in 3D SD-OCT images.
Methods: :
202 eyes of 58 healthy volunteers and 144 macular disease patients (38 MH, 117 ME, 116 ERM, and 52 AMD eyes) were scanned using SD-OCT (Cirrus HD-OCT; Carl Zeiss Meditec, Inc., Dublin, CA) with Macular Cube 200x200 scan protocol. A holistic data-driven approach, in which 3D OCT data were encoded using a rich set of dense, spatially-distributed features, is proposed. Machine learning algorithms were used to identify the most discriminative features automatically from our training data. For ground truth, one ophthalmologist labeled each scan with the pathologies it contained. A two-class non-linear support vector machine (SVM) was trained using leave-one-patient-out cross validation. For performance evaluation, areas under the receiver operating characteristics curve (AUC) were computed for classifications of healthy and four pathologies separately.
Results: :
We tested our method on 343 OCT scans. The AUC was 0.9714, 0.7862, 0.9338, 0.7939, 0.8325, for determining normal macula and the presence of MH, ME, ERM, and AMD, respectively.
Conclusions: :
We have developed an automated method for macular pathology diagnosis in OCT images and have obtained promising results, particularly in the cases of identifying the normal macula and identifying macular edema (AUC >0.9 for both). Our results demonstrate that the proposed holistic image representation combined with a data-driven learning framework can identify effective features without relying on a potentially error-prone segmentation module.
Keywords: image processing • imaging/image analysis: clinical • imaging methods (CT, FA, ICG, MRI, OCT, RTA, SLO, ultrasound)