Pseudo-nitzschia Cell Counter — YOLO11s-seg

A YOLO11s instance segmentation model fine-tuned to detect and count individual Pseudo-nitzschia cells within chain colonies in Imaging FlowCytobot (IFCB) images.

Try it out: Interactive demo on HF Spaces

Performance

Metric	Box	Mask
Precision	0.983	0.983
Recall	0.965	0.965
mAP50	0.976	0.973
mAP50-95	0.706	0.659

Trained for 100 epochs on Apple MPS (M-series GPU). Training time: ~1.5 hours.

Training Curves

Confusion Matrix

Quick Start

from ultralytics import YOLO
from huggingface_hub import hf_hub_download

# Download model
model_path = hf_hub_download("patcdaniel/pseudo-nitzschia-yolo11s-seg", "best.pt")
model = YOLO(model_path)

# Run inference
results = model("image.png", conf=0.25)

# Count cells
for r in results:
    n_cells = len(r.masks) if r.masks is not None else 0
    print(f"Detected {n_cells} cells")

Training

1. Download the dataset

huggingface-cli download patcdaniel/pseudo-nitzschia-cell-segmentation \
    --repo-type dataset --local-dir ./dataset

2. Train

python scripts/train_yolo.py \
    --config configs/cell_seg.yaml \
    --model yolo11s-seg.pt \
    --epochs 100 \
    --batch 24 \
    --name pn_cell_seg_yolo11s

3. Evaluate

python scripts/inference.py \
    --weights runs/segment/pn_cell_seg_yolo11s/weights/best.pt \
    --source dataset/data/images/ \
    --conf 0.25 \
    --save_viz

Model Details

Architecture: YOLO11s-seg (Ultralytics)
Base weights: yolo11s-seg.pt (COCO pre-trained)
Input size: 640x640
Classes: 1 (Pseudo_nitzschia_cell)
Parameters: ~11.8M
Model size: 20 MB

Training Hyperparameters

Parameter	Value
Epochs	100
Batch size	24
Image size	640
Optimizer	auto (SGD)
Initial LR	0.01
Final LR factor	0.01
Weight decay	0.0005
Warmup epochs	3.0
Patience (early stopping)	30
IoU threshold (NMS)	0.7
Augmentation	hsv, translate, scale, fliplr, erasing

Repository Structure

├── best.pt              # Model weights
├── configs/
│   └── cell_seg.yaml    # Dataset config (relative paths)
├── scripts/
│   ├── train_yolo.py    # Training script
│   ├── inference.py     # Inference & cell counting
│   ├── prepare_dataset.py
│   ├── coco_to_yolo_seg.py
│   ├── utils.py
│   ├── sample_inference.py
│   └── run_inference_on_dataset.py
├── results/             # Training metrics & curves
├── args.yaml            # Full training arguments
└── requirements.txt

Dataset

Trained on patcdaniel/pseudo-nitzschia-cell-segmentation — 302 annotated IFCB images from IFCB104 (Santa Cruz Municipal Wharf).

Citation

@model{pseudo_nitzschia_yolo11s_seg,
  author = {Daniel, Patrick C.},
  title = {Pseudo-nitzschia Cell Counter — YOLO11s-seg},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/patcdaniel/pseudo-nitzschia-yolo11s-seg}
}

Downloads last month: 5

patcdaniel
/

pseudo-nitzschia-yolo11s-seg