Artificial intelligence for classifying uncertain images by humans in determining choroidal vascular running pattern
Shozo Sonoda1,2, Hideki Shiihara1, Hiroto Terasaki1
1Department of Ophthalmology, Kagoshima University Graduate School of Medical and Dental Sciences, Kagoshima, Japan.
Plos One
|May 14, 2021
Summary
Artificial intelligence (AI) accurately classifies choroidal vascular patterns in pachychoroid diseases, matching human expert performance. This AI approach offers a reproducible method for analyzing these subtle vascular changes in eye imaging.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Abnormal choroidal vessel patterns are noted in pachychoroid diseases.
- Clinical assessment of these patterns lacks consistent reproducibility.
- Objective tools are needed for reliable evaluation.
Purpose of the Study:
- To compare the diagnostic concordance of choroidal vessel running patterns between AI and experienced clinicians.
- To evaluate the accuracy and reproducibility of AI in classifying these patterns.
Main Methods:
- Choroidal vessel patterns in 413 eyes were classified as symmetrical or asymmetrical.
- Three supervised machine learning models (SVM, Xception, Random Forest) were trained on clinician-labeled data.
- Confidence scores were used to compare AI certainty with human rater agreement.
Main Results:
- AI models achieved an area under the curve > 0.94 for pattern classification.
- The Random Forest model demonstrated the highest accuracy.
- AI agreement with humans showed higher certainty and lower uncertainty rates compared to disagreements.
Conclusions:
- AI algorithms can reliably classify choroidal vascular running patterns.
- AI demonstrates comparable accuracy and reproducibility to human experts.
- AI offers a potential solution for objective and consistent analysis of these vascular abnormalities.
