Related Experiment Video
Updated: Jan 18, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
Two-Step Semiautomated Classification of Choroidal Metastases on MRI: Orbit Localization via Bounding Boxes Followed
Jeffrey S Shi1, Bala McRae-Posani2, Sofia Haque2
1From the Department of Radiology (J.S.S., B.M.-P., S.H., A.H., J.S.), Memorial Sloan Kettering Cancer Center, New York, New York jss4002@med.cornell.edu.
AJNR. American Journal of Neuroradiology
|September 9, 2025
Summary
This study introduces an AI-powered deep learning framework to detect choroidal metastases in brain MRI scans. The novel approach accurately identifies these rare, often missed tumors, improving diagnostic capabilities.
Area of Science:
- Ophthalmology
- Radiology
- Artificial Intelligence
Background:
- Choroidal metastases are rare and often missed on brain MRI due to their small size and peripheral location.
- Early detection of choroidal metastases is crucial for effective patient treatment and management.
Purpose of the Study:
- To develop and validate an artificial intelligence (AI) framework for improved detection of choroidal metastases in brain MRI scans.
- To distinguish between normal orbits and those with choroidal metastases using a novel deep learning approach.
Main Methods:
- A hierarchical deep learning framework involving sequential orbit localization (YOLOv5) and binary classification was developed.
- A localization network was trained on 386 T2-weighted brain MRI slices, followed by a classifier trained on 66 patient MRIs (33 normal, 33 with metastases).
- Data-efficient evolutionary strategies were employed to address the challenge of a small dataset, mitigating overfitting and underfitting.
Main Results:
- The orbit localization model achieved 100% accuracy in identifying globes.
- The classification model, using evolutionary strategies-trained convolutional neural network (CNN), demonstrated a high area under the curve (AUC) of 0.93.
- The classifier achieved 100% sensitivity and 87% specificity for detecting choroidal metastases.
Conclusions:
- The developed semi-automated pipeline effectively detects small, easily overlooked choroidal metastases on brain MRI.
- This AI-driven approach shows significant clinical relevance for improving the diagnosis of choroidal metastases.
- Sequential localization and classification offer a promising strategy for identifying subtle lesions in medical imaging.
