Related Experiment Video
Updated: Jan 13, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
Uveal Pigmented Lesion Classification, Detection, and Segmentation: A Comparative Analysis of Machine Learning Tasks.
Virginia Tasso1,2, Sanjay Ganesh3, Sabrina Iddir3
1Department of Biomedical Engineering, University of Illinois Chicago, Chicago, IL, USA.
Translational Vision Science & Technology
|October 28, 2025
Summary
This study compares computer vision models for detecting uveal melanoma (UM). Classification models are most effective for UM detection with limited data, offering a valuable tool for regions with scarce resources.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Uveal melanoma (UM) is the most common adult intraocular malignancy with high metastatic potential and poor prognosis.
- Current methods for screening melanocytic choroidal tumors have limitations, especially in areas lacking specialized ocular oncologists.
Purpose of the Study:
- To propose a reference framework for future uveal melanoma detection research.
- To highlight the trade-offs between different computer vision (CV) approaches for UM detection based on data availability, resources, annotation effort, and clinical applicability.
Main Methods:
- Three CV models (classification, detection, segmentation) were developed using a ResNet-50 backbone on 864 Optos images of UM and other choroidal lesions.
- Performance was evaluated using AUC for classification, F1 score for detection, and Dice score for segmentation.
- An ablation study assessed robustness to data scarcity, with Grad-CAM and confusion matrices used for interpretability.
Main Results:
- Classification, detection, and segmentation models achieved AUC scores of 94%, 93%, and 95%, respectively.
- Segmentation yielded the highest F1 score (89%) and a Dice score of 75%.
- The classification model demonstrated superior performance in low-data scenarios.
Conclusions:
- In high-resource settings (100+ images), all CV models performed comparably.
- In low-resource settings (<70 images), the classification model significantly outperformed detection and segmentation models.
- Simpler classification models may provide greater utility for UM detection tasks when data resources are limited.
