Abstract
Purpose :
Machine learning algorithms are useful and efficient at interpreting medical images and segmenting anatomies. Here we present an approach that goes one step further by gaining scene understanding using cutting-edge machine learning techniques. Our method reliably detects anatomies of the anterior segment of the eye in OCT B-scans and implicitly understands the location of acquisition.
Methods :
The utilized neural network architecture in our work is U-net, a convolutional classifier without any fully connected layers, with multiple modifications. Batch renormalization is introduced to increase training efficiency. Squeeze and excitation layers are added to improve interdependencies between channels. Dilated convolutions are also used to increase the receptive field of the network. Our design emphasizes on scene understanding to accurately learn the correct position of anatomies relative to each other. The ADAM algorithm is used for training with cross entropy loss as cost function. The neural network classifies input image pixels as one of cornea, sclera, iris or background classes. A spectral domain OCT system is used for data acquisition. An automated method is employed to capture random OCT B-scans with varying parameters such as size, scale, location and gain from ex-vivo porcine eyes as training dataset. Multiple images are captured from each location as a form of data augmentation. Annotation of the anatomies in the acquired dataset is initialized by multiple automated algorithms and then manually refined.
Results :
In total 7503 training images and 1136 validation images are used. The network achieves an accuracy of 95.62% pixel classification over all classes and the entire validation dataset. Inference on a 1024×1024 OCT B-scan takes about 50 milliseconds.
Conclusions :
We have presented a reliable method using machine learning for real-time OCT anatomy classification with acceptable accuracy. The algorithm succeeds in segmentation independent of the input image size, scale or location. However, to train a neural network effectively for clinical use, gathering large datasets of human subjects or domain adaptation is crucial.
This abstract was presented at the 2019 ARVO Imaging in the Eye Conference, held in Vancouver, Canada, April 26-27, 2019.