| --- |
| title: DRIL OCT Classification |
| colorFrom: blue |
| colorTo: indigo |
| sdk: gradio |
| sdk_version: "5.29.0" |
| python_version: 3.11 |
| app_file: app.py |
| pinned: false |
| license: mit |
| --- |
| |
| # DRIL OCT Classification Demo |
|
|
| This Space provides an interactive demo for **Diabetic Retinopathy-related Inner Layer (DRIL)** detection from OCT B-scan images using fine-tuned deep learning models trained in the DRIL Classification Benchmark. |
|
|
| ## What is DRIL? |
|
|
| DRIL (Disruption of Retinal Inner Layers) refers to the loss of distinction between the inner plexiform, inner nuclear, and outer plexiform layers on OCT imaging. It is a recognized prognostic biomarker in diabetic macular edema associated with poor visual outcomes after treatment. |
|
|
| ## Models Available |
|
|
| | Model | Architecture | AUC (CV) | |
| |---|---|---| |
| | RETFound (Moderate FT) | Vision Transformer | Reported in paper | |
| | RETFound (Conservative FT) | Vision Transformer | Reported in paper | |
| | DenseNet-121 | CNN | Reported in paper | |
| | EfficientNet-B0 | CNN | Reported in paper | |
|
|
| ## How to Use |
|
|
| 1. Upload a fovea-centered OCT B-scan image (JPEG or PNG). |
| 2. Select the model you wish to use. |
| 3. Click "Classify" to get the DRIL probability and predicted label. |
|
|
| ## Notes |
|
|
| - The model was trained on a private OCT dataset (429 DRIL, 394 No-DRIL cases). |
| - Input images should be macular OCT B-scans. The model performs best on images similar to the training distribution. |
| - Test-time augmentation (TTA) is applied for more robust predictions. |
| - This demo is for research purposes only and should not be used for clinical decision-making. |
|
|
| ## Source Code and Weights |
|
|
| - GitHub: https://github.com/ai-research-2025/dril-classification-benchmark |
| - Model weights: https://drive.google.com/drive/folders/1MaenUEydngBCDaa-WqbwX8cGv5EzpYR_ |
|
|
