Related Experiment Video
Updated: Jun 18, 2025
05:46
Implantation and Evaluation of Melanoma in the Murine Choroid via Optical Coherence Tomography
Published on: December 2, 2022
1.7K
Using Deep Learning to Distinguish Highly Malignant Uveal Melanoma from Benign Choroidal Nevi
Laura Hoffmann1, Constance B Runkel1, Steffen Künzel1
1Department of Ophthalmology, Charité University Hospital Berlin, 12203 Berlin, Germany.
Journal of Clinical Medicine
|July 27, 2024
Summary
Deep learning software shows promise in identifying malignant choroidal lesions from fundus photos. This technology can aid in efficient pre-stratification of eye conditions.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Choroidal melanocytic lesions require accurate malignancy assessment.
- Deep learning (DL) offers potential for analyzing fundus photographs.
- Human-machine interaction (HMI) can enhance DL diagnostic capabilities.
Purpose of the Study:
- To evaluate DL software with HMI for choroidal lesion malignancy detection.
- To assess the performance of DL models using color fundus photographs (CFPs).
- To determine the software's utility in pre-stratifying lesion types.
Main Methods:
- Trained DL models on 762 CFPs of benign nevi, untreated, and irradiated melanomas.
- Utilized retinal specialists and multimodal imaging for reference standards.
- Evaluated trinary and binary classification models on 100 independent images.
Main Results:
- Achieved 84.8% accuracy for multi-class and 90.9% for binary classification.
- Demonstrated high recall (0.85–0.90) and specificity (0.91–0.92).
- Binary classification with a single imaging modality reached 95.8% accuracy (AUC 0.99).
Conclusions:
- DL models show strong performance in differentiating choroidal lesion malignancy.
- The software is a promising tool for resource-efficient pre-stratification.
- Potential for cost-effective preliminary assessment of ocular lesions.
