Deep learning architecture "LightOCT" for diagnostic decision support using optical coherence tomography images of biological samples

Abstract

Optical coherence tomography (OCT) is being increasingly adopted as a label-free and non-invasive technique for biomedical applications such as cancer and ocular disease diagnosis. Diagnostic information for these tissues is manifest in textural and geometric features of the OCT images, which are used by human expertise to interpret and triage. However, it suffers delays due to the long process of the conventional diagnostic procedure and shortage of human expertise. Here, a custom deep learning architecture, LightOCT, is proposed for the classification of OCT images into diagnostically relevant classes. LightOCT is a convolutional neural network with only two convolutional layers and a fully connected layer, but it is shown to provide excellent training and test results for diverse OCT image datasets. We show that LightOCT provides 98.9% accuracy in classifying 44 normal and 44 malignant (invasive ductal carcinoma) breast tissue volumetric OCT images. Also, >96% accuracy in classifying public datasets of ocular OCT images as normal, age-related macular degeneration and diabetic macular edema. Additionally, we show ∼96% test accuracy for classifying retinal images as belonging to choroidal neovascularization, diabetic macular edema, drusen, and normal samples on a large public dataset of more than 100,000 images. The performance of the architecture is compared with transfer learning based deep neural networks. Through this, we show that LightOCT can provide significant diagnostic support for a variety of OCT images with sufficient training and minimal hyper-parameter tuning. The trained LightOCT networks for the three-classification problem will be released online to support transfer learning on other datasets.

Fig. 1.

Micro electro mechanical system-vertical cavity surface emitting laser (MEMS-VCSEL) based swept source optical coherence tomography (SS-OCT) (OCS1310V1 - 1300 nm, Thorlabs) system used for imaging normal and cancer (invasive ductal carcinoma (IDC)) breast tissues (a) in the AIIMS dataset. Two illustrative examples of B-scan image (XZ) of normal and IDC breast tissues (b).

Fig. 2.

LightOCT architecture for the classification of different OCT images. LightOCT is used for classification of AIIMS dataset (normal and cancer breast tissue), Srinivasan dataset (normal, age-related macular degeneration and diabetic macular edema) and Zhang dataset (choroidal neovascularization, diabetic macular edema, drusen and normal samples). It features two convolutional layers and one fully connected layer. The hyper-parameters, N1, K1, N2 and K2 of the convolutional layers of LightOCT are tunable where N shows the size of kernel and K represents number of feature maps.

Fig. 3.

Illustration of the critical steps of LightOCT for normal and IDC cancer breast tissue OCT image of AIIMS dataset is shown here. Total 8 and 32 kernels are shown in the first and second convolution layer, respectively. The net effect of texture features (the first convolutional layer), local-cross feature patterns (the second convolutional layer), and cross-spatial cross-feature patterns (the fully connected layer) is evident in the outputs of the fully connected layer. The result of rectifier linear units and intermediate layers are not shown here for simplicity.

Fig. 4.

Kernels learnt by LightOCT for three different values of N1 are shown here for the AIIMS dataset. The kernels are resized 5 times with smoothing for the ease of visualization. These represent the texture features identified by the first convolutional layer.

Fig. 5.

The training and performance characteristics of LightOCT for normal and cancer breast tissue is shown here: (a) Loss curve, (b) accuracy curve and (c) receiver operating characteristics (ROC) curve.

Fig. 6.

Class activation map and prediction scores of normal/cancerous (AIIMS datasets) and ocular disease (Srinivasan and Zhang datasets) OCT images. (a-i): OCT images of three different datasets. The activation map is shown for LightOCT(a₁-i₁), after second convolutional layer of Inceptions-V3 (a₂-i₂) and just before the fully connected layer of Inception-V3 (a₃-i₃) network.