OR WAIT null SECS
Connor Iapoce is an assistant editor for HCPLive and joined the MJH Life Sciences team in April 2021. He graduated from The College of New Jersey with a degree in Journalism and Professional Writing. He enjoys listening to records, going to concerts, and playing with his cat Squish. You can reach him at email@example.com.
Anomaly detection techniques useful in identifying images with and without referable diabetic retinopathy when abnormal data were not available for training of retinal diagnostic systems.
In order to expand on a lack of data available for deep learning system (DLS) training for certain types of retinal diseases, a recent study explored the use of anomaly detection to identify common or rare ophthalmic diseases.
Led by Neil M. Bressler, MD, Johns Hopkins University School of Medicine, a team of investigators examined the application of anomaly detection to retinal diseases, finding it could be useful for diabetic retinopathy screening with potential for a broader application.
The study investigated a total of 16 different variants of anomaly detectors, including 12 generative and 4 discriminative detectors. Bressler and colleagues noted the 2 types of DLSs included discriminative networks performing discriminative tasks such as classification and segmentation, while generative networks generate realistic synthetic data.
Each anomaly detection method followed a 2-step embed-and-detect approach. Data were transformed to obtain new embedding and then, the new embedding was used to detect anomalies using a classic anomaly detection machine learning method.
Additionally, the study used the public EyePACs data set with 88,692 fundi of 44,346 individuals designed to be balanced across race and sex. Each image had a corresponding label from 0 - 4, representing different severity of diabetic retinopathy.
Further, all anomaly detection algorithms were trained only with normal retinas and challenged to identify previously unseen diseased retinas. The anomaly detectors were evaluated by commonly accepted performance metrics, including area under the receiver operating characteristic curve, F1 score, and accuracy. The analysis took place from September 2019 - September 2020.
Data show the combination of using Deep InfoMax for embedding and 1-class support vector machine as anomaly detection achieved the best performance across all criteria. The area under the receiver operating characteristic was the best performing at 0.808 (95% CI, 0.789 - 0.827), with an accuracy of 73.90% (95% CI, 71.98% - 75.82%), and F1 of 0.755.
Then, among generative anomaly detection methods, InfoStyleGAN embedding achieved the best performance across all metrics, with the area under the receiver operating characteristic of 0,667 (95% CI, 0.644 - 0.691). Overall, investigators found discriminative embedding systems outperformed generative ones.
Overall, investigators noted that the findings suggested when abnormal data, such as referable diabetic retinopathy, were not available for training of retinal diagnostic systems, anomaly detection techniques showed usefulness in identifying images with and without referable diabetic retinopathy.
“This investigation suggests anomaly detection could be useful for diabetic retinopathy screening for retinal diseases, and its use could be pursued in broader applications, eg, detecting rare diseases or new presentations of common retinal diseases,” Bressler and colleagues wrote.
The study, “Detecting Anomalies in Retinal Diseases Using Generative, Discriminative, and Self-supervised Deep Learning,” was published in JAMA Ophthalmology.