Neal S. Shah*, Aniket Ramshekar*, Bright Asare-Bediako, Morgan P. Tankersley, Heng-Chiao Huang, Shreya Beri, Eric Kunz, Aaron Y. Lee, M. Elizabeth Hartnett
Byers Eye Institute Department of Ophthalmology, Stanford University School of Medicine, Stanford, CA, USA
Segmentation, Retinal Flatmount, Oxygen-Induced Retinopathy, OIR, Mouse, Rat, Intravitreal Neovascularization, Avascular Area
This is a repo to download model weight checkpoints for our model that can be found on the MONAI model zoo. This model performs automated segmentation of oxygen-induced retinopathy (OIR) retinal flatmount images into three regions: total retina (TR), intravitreal neovascularization (IVNV), and avascular area (AVA).
The architecture is a multi-task Attention U-Net with a ConvNeXt-Tiny encoder [1] and deep supervision (~8.7M trainable parameters).
For inference, the final release uses an ensemble of 5 cross-validation models, with test-time augmentation and per-class thresholding, to improve robustness across mouse and rat OIR images.
Model development used three datasets:
- Rat IVNV pretraining dataset: 72 rat OIR flatmount images with IVNV-only annotations (used in intermediate Stage 2 training).
- Final development dataset: 345 annotated images total (267 mouse, 78 rat), including:
- 127 expert human-annotated images (49 mouse, 78 rat)
- 218 curated open-source mouse images [2] with reviewed masks generated from a prior published model [3]
- Independent test dataset: 37 images (18 mouse OIR, 19 rat OIR), held out from training/validation/model selection.
For final model development, a modified 5-fold cross-validation strategy was used, with expert-annotated images serving as fold-level validation references and curated open-source mouse images used in training only.
Input retinal flatmount images were converted to grayscale, resized to 512×512, and intensity-normalized.
During training, joint image-mask augmentation was applied using random horizontal/vertical flips, random rotations (up to 180 degrees), brightness/contrast perturbation, CLAHE, Gaussian noise, elastic/grid/optical distortions, coarse dropout, motion blur, and random gamma adjustments.
Dice agreement between model masks and human consensus masks was high for total retina (TR) and AVA, and moderate for IVNV in both species:
- Rat: TR Dice=0.983, AVA Dice=0.924, IVNV Dice=0.612
- Mouse: TR Dice=0.975, AVA Dice=0.912, IVNV Dice=0.601
At the metric level, the deep learning model showed strong correlation with the mean of three graders for rat percent AVA (r=0.979) and rat percent IVNV (r=0.943). In mouse OIR, correlation was strong for percent AVA (r=0.957) but weak for percent IVNV (r=0.265), likely due to high inter-grader variability for mouse IVNV scoring.
(For full analysis please refer to the manuscript.)
This model was trained on an Apple M2 pro 16GB Macbook. 5-fold cross-validation was run sequentially with batch size 4 and a maximum of 120 epochs per fold. The folds ran for 86, 88, 120, 61, and 80 epochs, with total training time of approximately 24 hours.
Model checkpoints are hosted externally and linked through large_files.yml (not committed directly in the repo due to file size limits).
- Liu Z, Mao H, Wu CY, Feichtenhofer C, Darrell T, Xie S. A ConvNet for the 2020s. 2022:11966-11976.
- Marra KV, Chen JS, Robles-Holmes HK, et al. Development of an Open-Source Dataset of Flat-Mounted Images for the Murine Oxygen-Induced Retinopathy Model of Ischemic Retinopathy. Transl Vis Sci Technol. Dec 2 2024;13(12):4.
- Xiao S, Bucher F, Wu Y, et al. Fully automated, deep learning segmentation of oxygen-induced retinopathy images. JCI Insight. Dec 21 2017;2(24)doi:10.1172/jci.insight.97585