File size: 3,028 Bytes
fd0e3e0 109c1a2 fd0e3e0 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 | ---
license: mit
tags:
- object-detection
- instance-segmentation
- medical-imaging
- microbiology
- antibiotic-susceptibility-testing
library_name: pytorch
pipeline_tag: object-detection
---
# ZoneVision — Inhibition-Zone Detection for AST
Automated inhibition-zone (antibiotic halo) detection and quantitative measurement on 96-well plate photographs for antibiotic susceptibility testing (AST).
## Model Weights
| File | Size | Architecture | Purpose |
|------|------|-------------|---------|
| `rfdetr_seg_small_best.pth` | 128 MB | RF-DETR-Seg-Small (33.4M params) | End-to-end zone instance segmentation |
| `sam3.pt` | 3.2 GB | SAM3 | Optional mask refinement within detected ROIs |
| `yolo26n.pt` | 5.3 MB | YOLO26n | Pretrained backbone for plate geometry estimation |
| `yolo26n-seg.pt` | 6.4 MB | YOLO26n-seg | YOLO segmentation model (alternative detector) |
## Performance
| Metric | Value |
|--------|-------|
| F1 Score | 0.952 |
| Precision | 0.973 |
| Recall | 0.931 |
| Mean IoU | 0.896 |
| Diameter MAE | 0.234 mm (3.08% relative) |
| Pearson r (diameter) | 0.973 |
Evaluated on 11 plate photos with 233 manually annotated inhibition zones.
## Pipeline
1. **Plate geometry** — YOLO26n + Hough Circles detect the 96-well grid; estimate px/mm from 9.0 mm well pitch
2. **Zone segmentation** — RF-DETR-Seg-Small produces per-zone masks
3. **Mask refinement** (optional) — SAM3 refines boundaries
4. **Measurement** — Pixel-to-mm conversion, diameter/area extraction, QC flags
5. **Output** — CSV with per-well phenotypes, overlay images, binary masks
## Quick Start
```bash
# Install
pip install -e .
# Download weights
hf download logichenry/ZoneVision --local-dir weights/
# Run inference
python scripts/run_pipeline.py \
--input path/to/plate_photos/ \
--output outputs/ \
--config configs/config.yaml \
--detector rfdetr
```
## Training
The RF-DETR model was trained on 233 annotated inhibition zones across 11 plate photos in COCO format. See the [GitHub repo](https://github.com/SmartisanNaive/ZoneVision) for training scripts and dataset preparation tools.
## Intended Use
- Automated measurement of inhibition zones in antibiotic susceptibility testing
- High-throughput screening of antimicrobial peptide libraries on 96-well plates
- Quantitative phenotyping for lanthipeptide or bacteriocin activity assays
## Limitations
- Designed for color photographs of 96-well plates; may not generalize to other formats
- SAM3 refinement requires ~3.2 GB VRAM; can be disabled for resource-constrained environments
- Best performance on plates with clear zone boundaries; heavily overlapping zones may reduce accuracy
## Citation
```bibtex
@article{zonevision2026,
title={Automated Inhibition-Zone Detection for Antibiotic Susceptibility Testing Using Cascade Vision},
author={Baice},
journal={Chinese Journal of Biotechnology},
year={2026}
}
```
## License
MIT License. See [LICENSE](https://github.com/SmartisanNaive/ZoneVision/blob/main/LICENSE).
|