Automated Detection of Epiretinal Membranes in OCT Scans

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Apr 20 2024

In a paper published in the journal Scientific Reports, researchers developed a deep neural network (DNN) to automatically detect and classify epiretinal membranes (ERMs) in optical coherence tomography (OCT) scans of the macula region by utilizing over 11,000 images.

The DNN ensemble achieved high accuracy in detecting no ERM, small ERMs, and large ERMs, with small ERMs presenting the greatest challenge while employing t-SNE analysis for dimensional reduction and saliency maps to highlight ERMs, even in complex cases. This study demonstrated the DNN's potential for accurately detecting and grading ERMs, offering promise for future screening programs or decision-support systems.

Background

Past studies have explored the automated detection of ERMs using funduscopy, retinal fundus images, and OCT. However, these studies often relied on limited datasets or focused on advanced stages of the disease. However, such approaches have typically encountered challenges in detecting early and low-grade stages of ERMs.

Moreover, the reliance on small datasets may limit the generalizability of findings to broader clinical populations. Furthermore, the limited datasets used in previous studies may not adequately represent the diverse manifestations of ERMs encountered in clinical practice. It could hinder the robustness of automated detection systems in real-world scenarios.

Advanced ERM Detection

The dataset comprised 624 OCT volume scans obtained from 624 eyes of 461 patients. These scans yielded 11,061 individual brightness scans, also known as B-scans. Most patients in the dataset were Caucasian, and the OCT images were collected using Heidelberg Spectralis OCT. Most scans were standardized horizontal fovea-centered volume scans containing 25 cross-sections, with additional single scans through the fovea in approximately 3% of cases.

An ERM was defined as a hyperreflective membrane on the retina's inner surface, and a retina specialist graded scans according to the presence and size of ERMs, with exclusion criteria applied to ensure data quality, including secondary ERMs, high myopia, and poor-quality OCT images.

This study employed deep learning (DL) models based on the residual network with 50 layers (ResNet50) and InceptionV3 architectures, which were pre-trained on ImageNet and fine-tuned for ERM classification tasks. The architecture of these models involved a combination of max pooling and average pooling in the convolutional stack, followed by dense layers with batch normalization and rectified linear unit (ReLU) activation.

Experimenters employed random oversampling to ensure a more balanced representation of different classes by training models using stochastic gradient descent with Nesterov's accelerated gradients to counteract class imbalance in the data. Additionally, a mixup was applied as a data augmentation technique to enhance model generalization.

Ensemble learning further enhanced ERM detection performance, especially in challenging scenarios. Ensembles accurately detected and classified ERMs according to their size. Moreover, uncertainty estimates provided insights into the reliability of model predictions, with well-calibrated uncertainty observed on test data. Utilizing t-stochastic neighborhood embeddings (t-SNE) helped gain insights into the feature representations learned by the models. This non-linear dimensionality reduction method allowed for interpreting high-dimensional data in low dimensions, providing valuable information about the ensemble-based representations.

Finally, saliency maps were generated using the guided backprop algorithm to enhance the interpretability of the models' decisions. These maps highlighted important areas of interest on OCT scans, aiding in detecting ERMs, even small ones not located in the fovea. Overall, the study integrated advanced DL and image analysis techniques to develop a robust automated ERM detection and classification system with potential applications in clinical practice.

Training Process Enhancements

During the training process, the analysts employed a technique known as mixup along with standard data augmentation operations to enhance ERM detection performance. Interestingly, the effectiveness of the DNNs improved with increased mixing and longer training durations, particularly in scenarios presenting more clinical relevance and complexity.

The DNNs achieved accurate detection of ERMs and could also determine the size of the detected ERMs with high precision. Notably, the best ensemble model obtained a 3-way classification accuracy of 89.33% and highly differentiated between normal B-scans and those with small or large ERMs. However, small ERMs posed a greater challenge for the DNNs than larger ones, likely due to the inherent difficulty in identifying such fine structures. Ablation of ensembling along with mixup adversely affected the detection rate of small ERMs, emphasizing the importance of these techniques in achieving robust performance.

The uncertainty estimates provided by the ensemble model were well-calibrated on the test data, indicating their potential utility in providing valuable information about the model's performance. Saliency maps highlighted important areas of interest, particularly on the retina's inner surface, aiding in detecting even small ERMs. These maps also revealed that associated retinal changes and entities were not crucial for the DNN's decision-making process, further highlighting the model's ability to focus on relevant features.

Additionally, the study explored the representation learned by the DNNs using t-SNE, which accurately ordered discrete ERM classes according to their severity. Pairing t-SNE coordinates with predictive uncertainty provided insights into the model's decision boundaries, with high uncertainty observed at transitions between ERM-negative and positive cases. These findings demonstrate the effectiveness of DNNs in detecting and classifying ERMs, with implications for improving diagnostic accuracy in ophthalmological settings.

Conclusion

To sum up, the study demonstrated the efficacy of DL models, particularly ResNet50 and InceptionV3 architectures, in accurately detecting and classifying ERMs from OCT scans. Leveraging ensemble learning techniques improved performance, especially in challenging scenarios, while providing reliable uncertainty estimates.

Additionally, the interpretability of the models was enhanced through the generation of saliency maps, aiding in understanding the decision-making process. Overall, the approach showcased the potential of advanced DL methods in automated ERM detection and classification, offering valuable support to clinicians in diagnosis and treatment planning.