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 long process of conventional diagnostic procedure and shortage of human expertise. Here, a custom deep learning architecture, LightOCT, is proposed for classification of OCT images into diagnostically relevant classes. LightOCT is a convolutional neural network with only 2 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 > 99% accuracy in classifying 44 normal and 44 cancerous (invasive ductal carcinoma) breast tissues collected from 22 patients to perform the study. Also, ~99% 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. Through this, we show that LightOCT can provide simultaneous significant diagnostic support for variety of OCT images with sufficient training and very little hyper-parameter tuning.

Inflation of test accuracy due to data leakage in deep learning-based classification of OCT images

“…The results obtained on Kermany's original_v2 split are in accordance with reported accuracy [18][19][20] , where the differences in performance can be attributed to the deep learning model used and its optimization. However, results on the third version of the dataset were much lower compared to those in literature 10,21,22 , even when using the same model architecture, loss and optimizer 10 . Interestingly, Table 1 also shows that studies using multiple datasets and reporting different split strategies for each dataset 10,23,24 , report as high accuracy on the per-volume/subject split datasets as the one on the original_v2 or the per-image split datasets.…”

“…However, results on the third version of the dataset were much lower compared to those in literature 10,21,22 , even when using the same model architecture, loss and optimizer 10 . Interestingly, Table 1 also shows that studies using multiple datasets and reporting different split strategies for each dataset 10,23,24 , report as high accuracy on the per-volume/subject split datasets as the one on the original_v2 or the per-image split datasets. Moreover, the drop in MCC seen when using Kermany's original_v2 compared to the original_v3 split, highlights the inflation effect that leakage between training and testing data has on model performance, especially when the original_v2 shows to have 92% overlap between training and testing sets while original_v3 has none.…”

“…Model architecture and training strategy. The LightOCT model proposed by Butola et al 10 was used in this study. LightOCT is a custom, shallow and multi-purpose network for OCT image classification composed of a two-layer CNN encoder, and one fully connected layer with softmax activation as output layer.…”

“…The first and second convolutional layers have 8 and 32 convolutional filters, respectively. The kernel size of the filters in both layers is set to 5 × 5 and the output of each layer passes through a ReLU activation function 10 . A max-pooling operation is present between the first and the second convolutional layer that cuts the spatial dimension of the output of the first layer in half.…”

“…As for the case of many of the reviewed medical image analysis challenges 8 , one aspect that is sometimes missing or not well described is how the testing dataset is generated from the original pool of data. Moreover, there are examples where the preparation of the testing dataset was described, but its overlap with the training set was not considered [9][10][11][12][13] , undermining the reliability of the reported results. Focusing on deep-learning applications for OCT, depending on the acquisition set-up, volumes are usually acquired with micrometer resolution in the x, y and z directions in a restricted field of view, with tissue structures that are alike and affected by similar noise.…”