Automating Disease Detection in Retinal OCT Scans with RETFound
Author Information
Author(s): Du Katherine, Nair Atharv Ramesh, Shah Stavan, Gadari Adarsh, Vupparaboina Sharat Chandra, Bollepalli Sandeep Chandra, Sutharahan Shan, Sahel José-Alain, Jana Soumya, Chhablani Jay, Vupparaboina Kiran Kumar, Mesin Luca, Ohguro Hiroshi
Primary Institution: University of Pittsburgh Medical Center
Hypothesis
Can the RETFound model improve the detection of disease features in retinal OCT scans compared to traditional methods?
Conclusion
The RETFound model shows comparable performance to ResNet-50 in detecting disease features in retinal OCT scans, potentially improving diagnostic accuracy.
Supporting Evidence
- RETFound was trained on 1.6 million unlabeled retinal OCT images.
- The accuracy of diagnosing healthy versus diseased OCT B-scans was high across all model types.
- RETFound showed improved interpretability compared to traditional CNN models.
- External validation demonstrated high accuracy for both models on an independent dataset.
- RETFound requires less labeled data than traditional models.
- Specific features like drusen and PED were identified with higher accuracy.
- Inter-observer variance was evaluated to ensure labeling consistency.
- Automated algorithms can help standardize medical diagnoses.
Takeaway
Researchers used a computer program to help doctors find eye diseases in pictures of the eye, making it easier and faster to spot problems.
Methodology
The study involved training the RETFound model on 1770 labeled OCT B-scans and comparing its performance with the ResNet-50 model.
Potential Biases
Potential bias due to reliance on a single reviewer for labeling OCT scans.
Limitations
The study could not evaluate the models on certain features due to a lack of labeled scans and was limited to data from one OCT device.
Participant Demographics
110 unique eyes from individuals diagnosed with age-related macular degeneration (AMD).
Statistical Information
P-Value
0.05
Statistical Significance
p<0.05
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website