HydroMorph Ventricle SegFormer-B0
Ventricle segmentation model for Normal Pressure Hydrocephalus (NPH) Evans Index measurement.
Status: This model repository is a placeholder for the SegFormer-B0 fine-tuned on brain MRI ventricle segmentation data. The current HydroMorph pipeline uses a classical computer vision approach (adaptive thresholding + Otsu + connected-component analysis) which achieves 100% sensitivity at 7.5ms/image without requiring GPU inference.
Context: HydroMorph Pipeline
This model is part of the HydroMorph project — the first benchmark for NPH diagnostic AI, comparing classical CV vs vision-language models for automated Evans Index measurement.
Why Classical CV First?
| Approach | Sensitivity | Specificity | Speed | GPU Required |
|---|---|---|---|---|
| Classical CV (current) | 100% | 50% | 7.5 ms/image | No |
| SegFormer-B0 (planned) | TBD | TBD | ~50 ms/image | Yes |
| VLM Agent (GPT-4o) | 50% | 80% | 8 sec/image | No (API) |
| Cascade (CV → VLM) | 100% | ~80% | — | — |
The classical CV pipeline already achieves 100% sensitivity as a screening tool. The SegFormer model is planned as a potential replacement or complement, offering:
- Learned features vs hand-crafted thresholds
- Better generalization across imaging protocols
- Pixel-level segmentation maps for volumetric analysis
Architecture
- Base: nvidia/segformer-b0-finetuned-ade-512-512
- Task: Binary semantic segmentation (ventricle vs background)
- Input: 512×512 brain MRI (axial slices)
- Output: Binary mask + Evans Index computation from mask geometry
- Training data: mmrech/hydromorph-demo-segmentation-seed
Intended Use
- Research tool for automated ventriculomegaly screening
- Not for clinical diagnosis
- Part of the HydroMorph two-stage cascade pipeline
Links
- Demo: HydroMorph Space
- Dataset: mmrech/hydromorph-nph
- Code: github.com/mmrech/hydromorph
- Paper: Hydrocephalus 2026 World Congress, Athens, Greece
Citation
@inproceedings{rech2026hydromorph,
title={HydroMorph: AI-Assisted Automated Evans Index Measurement
Combining Classical Morphometry with Vision-Language Model
Self-Correction},
author={Rech, Matheus Machado},
booktitle={Hydrocephalus 2026 World Congress},
year={2026},
address={Athens, Greece}
}