Related Experiment Video
Updated: Jul 21, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Deep Learning for Localized Detection of Optic Disc Hemorrhages.
Aaron Brown1, Henry Cousins2, Clara Cousins3
1From the Department of Ophthalmology (A.B., K.E., A.H., A.F., L.B., K.Y., K.V., N.C., L.R.P.), Icahn School of Medicine at Mount Sinai, New York, New York, USA; Department of Ophthalmology (A.B., K.E., A.H., A.F., L.B., K.Y., K.V., N.C., L.R.P.), New York Eye and Ear Infirmary of Mount Sinai, New York, New York, USA.
An automated deep learning system effectively detects disc hemorrhages in optic disc photographs. This object detection model outperforms image classification and matches clinician performance, aiding in glaucoma diagnosis.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Disc hemorrhages are critical indicators of glaucoma progression.
- Accurate detection of disc hemorrhages is essential for timely diagnosis and management.
- Automated systems can potentially improve the efficiency and consistency of hemorrhage detection.
Purpose of the Study:
- To develop and evaluate a deep learning system for automated detection and localization of disc hemorrhages.
- To compare the performance of an object detection model against an image classification model.
- To assess the system's performance relative to expert clinicians.
Main Methods:
- Trained two deep learning models (object detection and image classification) on a dataset of 1562 optic disc photographs from 5 institutions.
- Utilized expert graders for initial image classification.
- Compared model performance using Area Under the Receiver Operating Characteristic Curve (AUC) and evaluated against two glaucoma specialists.
Main Results:
- The object detection model achieved a superior AUC of 0.936 compared to the image classification model's AUC of 0.845.
- At high specificity, the object detection model demonstrated 94.3% specificity and 70.0% sensitivity, comparable to expert clinicians.
- At high sensitivity, the model achieved 96.7% sensitivity and 73.3% specificity.
Conclusions:
- An autonomous object detection model is more effective than image classification for identifying disc hemorrhages.
- The developed deep learning system performs comparably to expert clinicians in detecting disc hemorrhages.
- This automated system shows promise for improving glaucoma diagnosis and monitoring.
