README.md

metadata

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

# 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 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

@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.