Animesh-Kr/oct-fluid-segmentation: v1.0.0 — Attention-Guided TransUNet for Retinal Fluid Segmentation

Attention-Guided TransUNet for Multi-Class Retinal Fluid Segmentation

What this release includes

• Complete training pipeline for IRF / SRF / PED segmentation across 4 OCT sources
• Dual AttentionTransUNetL V2L ensemble (EfficientNetV2L encoder, 127M params)
• Source-Adaptive BatchNorm for cross-scanner domain adaptation
• MC Dropout + inter-model disagreement dual uncertainty estimation
• UCUS — Uncertainty-Weighted Clinical Urgency Score (Monitor / Review / Urgent)
• Streamlit dashboard with live ONNX inference
• FastAPI inference endpoint
• INT8 quantisation (V2L 510MB → 132MB, 3.9× compression)

Results

| Metric | Value | |--------|-------| | V2L val Dice (mean ± std) | 0.784 ± 0.006 | | V2L IRF Dice | 0.916 ± 0.003 | | V2L SRF Dice | 0.856 ± 0.003 | | V2L PED Dice | 0.581 ± 0.018 | | Uncertainty ratio at disagreement | 1.34× (p=3.77e-05) | | SRF volume correlation | r=0.778 (p=6.33e-04) | | PED volume correlation | r=0.841 (p=8.64e-05) |

Model weights

All checkpoints hosted on HuggingFace: https://huggingface.co/animeshakr/oct-fluid-segmentation

Live demos

• Dashboard: https://huggingface.co/spaces/animeshakr/oct-fluid-segmentation
• API: https://huggingface.co/spaces/animeshakr/oct-fluid-segmentation-api
• Complete pipeline: https://huggingface.co/spaces/animeshakr/oct-complete-pipeline

Datasets

DUKE DME, AROI (Melinščak et al. 2021), UMN AMD/DME (Parhi, University of Minnesota)

Citation

@misc{kumar2026octseg, title={Attention-Guided TransUNet for Multi-Class Retinal Fluid Segmentation in OCT with MC Dropout Uncertainty Quantification}, author={Kumar, Animesh}, institution={Newcastle University}, year={2026} }