Related Experiment Video
Updated: Jan 13, 2026
09:53
Quantifying the Brain Metastatic Tumor Micro-Environment using an Organ-On-A Chip 3D Model, Machine Learning, and Confocal Tomography
Published on: August 16, 2020
7.6K
Benchmarking diffusion models against state-of-the-art architectures for OCT fluid biomarker segmentation
Katherine Du1, Utkarsh Doshi1, Benjamin DiCenzo1
1Department of Ophthalmology, University of Pittsburgh Medical Center, Pittsburgh, Pennsylvania, United States of America.
Plos One
|October 29, 2025
Summary
Deep learning models can segment retinal diseases from OCT scans. While diffusion models show promise, nnU-Net performed best for automated analysis of subretinal fluid, intraretinal fluid, and pigment epithelial detachment.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Retinal diseases cause significant vision loss, necessitating accurate diagnostic tools.
- Optical coherence tomography (OCT) is crucial for assessing retinal pathologies.
- Automated interpretation of OCT scans using deep learning can standardize disease assessment.
Purpose of the Study:
- To evaluate the performance of a diffusion model for segmenting subretinal fluid (SRF), intraretinal fluid (IRF), and pigment epithelial detachment (PED) in OCT scans.
- To compare the diffusion model's segmentation accuracy against other leading deep learning models.
- To assess the potential of automated segmentation for clinical applications in retinal disease management.
Main Methods:
- Manual segmentation of SRF, IRF, and PED in 269, 224, and 114 OCT scans, respectively, by three reviewers.
- Training and evaluation of diffusion model, Nested U-Net, nnU-Net, TransUNet, and SwinUNet using 5-fold cross-validation.
- Performance metrics included Dice coefficient, sensitivity, specificity, Pearson correlation coefficient, and R2.
Main Results:
- All evaluated models demonstrated high similarity to ground truth segmentations.
- The diffusion model exhibited relatively higher sensitivity compared to most other models.
- nnU-Net achieved the strongest overall performance, indicating superior accuracy in automated OCT analysis.
Conclusions:
- Diffusion models are capable of segmenting retinal pathologies comparably, even with limited annotated data.
- nnU-Net is identified as the most effective model for automated analysis of retinal OCT scans.
- These findings support the advancement of deep learning for standardized and efficient retinal disease diagnosis.
