Source-linked AI summary
Deep learning is effective for the classification of OCT images of normal versus Age-related Macular Degeneration
Cecilia S. Lee, Doug M. Baughman, Aaron Y. Lee
TL;DR
The paper examines whether deep learning can distinguish AMD from normal OCT images using data linked to electronic medical records. It reports effective classification, with performance improving when predictions are aggregated at the macular-scan and patient levels.
Problem
The paper addresses whether deep learning can distinguish AMD from normal OCT images, an application described as novel in ophthalmology.
Method
The study applies a deep learning model to OCT classification using training and validation images extracted from a large electronic medical record database.
Results
Deep learning effectively distinguished AMD from normal OCT images, with higher accuracy when probabilities were aggregated at the macular-scan and patient levels.
Takeaways & Limitations
The findings demonstrate the feasibility of applying deep learning to OCT classification in ophthalmology.
Takeaways & Limitations
Model training used images from only a single academic center.
Abstract
from arXiv · showhide
Objective: The advent of Electronic Medical Records (EMR) with large electronic imaging databases along with advances in deep neural networks with machine learning has provided a unique opportunity to achieve milestones in automated image analysis. Optical coherence tomography (OCT) is the most commonly obtained imaging modality in ophthalmology and represents a dense and rich dataset when combined with labels derived from the EMR. We sought to determine if deep learning could be utilized to distinguish normal OCT images from images from patients with Age-related Macular Degeneration (AMD). Methods: Automated extraction of an OCT imaging database was performed and linked to clinical endpoints from the EMR. OCT macula scans were obtained by Heidelberg Spectralis, and each OCT scan was linked to EMR clinical endpoints extracted from EPIC. The central 11 images were selected from each OCT scan of two cohorts of patients: normal and AMD. Cross-validation was performed using a random subset of patients. Area under receiver operator curves (auROC) were constructed at an independent image level, macular OCT level, and patient level. Results: Of an extraction of 2.6 million OCT images linked to clinical datapoints from the EMR, 52,690 normal and 48,312 AMD macular OCT images were selected. A deep neural network was trained to categorize images as either normal or AMD. At the image level, we achieved an auROC of 92.78% with an accuracy of 87.63%. At the macula level, we achieved an auROC of 93.83% with an accuracy of 88.98%. At a patient level, we achieved an auROC of 97.45% with an accuracy of 93.45%. Peak sensitivity and specificity with optimal cutoffs were 92.64% and 93.69% respectively. Conclusions: Deep learning techniques are effective for classifying OCT images. These findings have important implications in utilizing OCT in automated screening and computer aided diagnosis tools.
INTRODUCTION
OCT use and storage in large electronic databases create an opportunity for automated analysis, while deep learning can learn image features from training data. The study therefore linked OCT images with EMR data to develop an algorithm distinguishing AMD from normal OCT images.
- Motivation: OCT is widely used in ophthalmology, with images increasingly stored in large electronic databases that increase provider time and effort.
- Motivation: AMD-associated OCT findings differ from those of a normal retina, making automated identification potentially useful for directing clinicians to pathological regions.
- Deep learning: Deep learning trains many-layered neural networks to develop convolutional matrices from training data and has improved image classification and object detection.
- Problem: Deep learning often requires several tens of thousands of examples, limiting its application when large training sets are unavailable.
- Study objective: The study linked large OCT and EMR datasets to train a deep learning algorithm that distinguishes AMD from normal OCT images.
OCT and EMR Extraction
Macular OCT scans were automatically extracted from the Heidelberg Spectralis imaging database, with every image linked to EPIC clinical data by patient medical record number and encounter dates. Extracted clinical data included diagnoses, encounter dates, macular laser procedures, and intravitreal injections.
- OCT acquisition and extraction: Every image from each macular OCT scan was extracted from the 61-line raster macula scans.
- EMR linkage: OCT images were linked to clinical data in EPIC using patient medical record numbers and encounter dates.
- EMR linkage: EPIC Clarity tables supplied clinical diagnoses, encounter dates, macular laser procedures, and intravitreal injections.
Patient and Image Selection
Patients were classified as normal or AMD using predefined clinical criteria, with other macular pathology excluded. EMR-linked OCT images were processed using central foveal slices and patient-exclusive training and validation sets.
- Normal patients had no retinal ICD-9 diagnosis and better than 20/30 vision in both eyes throughout recorded UW clinical history.
- AMD patients had a retina-specialist ICD-9 AMD diagnosis, at least one intravitreal injection, and worse than 20/30 vision in the better-seeing eye.
- Patients with other macular pathology by ICD-9 code were excluded, while consecutive images from eligible patients were retained without image-quality exclusions.
- EMR labels were linked to OCT macular images, and protected health identifiers were removed.
- The central 11 images from each macular OCT set were independently labeled normal or AMD, histogram equalized, and down-sampled to 192x124.
- 20% of patients in each group formed the validation set, with mutually exclusive patients between validation and training sets and randomized training-image order.
Deep Learning Classification Model
A modified VGG16 convolutional neural network classified OCT images as normal or AMD, with probabilities aggregated for macular OCT and patient-level analyses. Performance was strongest at the patient level, and occlusion testing identified pathological OCT regions important for AMD categorization.
- Model interpretation: Occlusion testing identified pathological OCT regions as the most critical image areas for the network’s AMD categorization.A blank 20x20 pixel box was moved across each image, and the largest probability drop indicated the most important region.
- Macular OCT-level performance: 93.82% AUROC and 88.98% accuracy were achieved at the macular OCT level by averaging probabilities across images from the same OCT scan.Sensitivity and specificity were 85.41% and 93.82%, respectively.
- Patient-level performance: 93.45% accuracy, 83.82% sensitivity, and 96.40% specificity were achieved at the patient level by averaging image probabilities from the same patient.The patient-level AUROC was 97.46%, with optimal-cutoff sensitivity and specificity of 92.64% and 93.69%.
- Image-level performance: 92.77% AUROC and 87.63% accuracy were achieved at the individual-image level, with 84.63% sensitivity and 91.54% specificity.Optimal-cutoff sensitivity and specificity were 87.08% and 87.05%, respectively.
DISCUSSION
The study demonstrates that deep learning can distinguish AMD from normal OCT images, with improved performance when predictions are aggregated across macular scans and patients. Occlusion testing identified pathological regions as important features, while external generalizability remains unestablished.
- The deep learning neural network effectively distinguished AMD from normal OCT images.
- Aggregating probabilities across OCT macular scans and patients increased accuracy, mainly through improved sensitivity with more images.AMD may not affect the entire macula, allowing normal-appearing images within scans from AMD patients.
- The model used 21 neuron layers with state-of-the-art convolutional networking layers, extending deep learning applications to OCT classification.The authors report that training and validation images from a large EMR extraction had not previously been shown in this context.
- Occlusion testing showed that the network identified OCT regions corresponding to pathology and did not strongly depend on nasal high choroidal transmission.
- The model was trained on selected images from a single academic center, so its external generalizability is unknown.Future validation should include other institutions, OCT manufacturers, all images from macular scans, and additional diagnoses.
- Automated OCT classification could support retinal pathology screening and help clinicians identify concerning macular images for diagnosis and treatment.The inference step is computationally inexpensive compared with training and can run on standard computers.