esab/pbc-cell-classifier

ResNet-18 Peripheral Blood Cell Classifier

Model Description

This is a ResNet-18 model fine-tuned for peripheral blood cell (PBC) classification using fastai. The model can classify blood cell images into 8 different cell types with 98.07% validation accuracy.

Model Details

• Model Type: ResNet-18 with transfer learning
• Framework: fastai (version <2.8.0)
• Task: Image Classification
• Dataset: Peripheral Blood Cell (PBC) dataset
• Classes: 8 cell types
• Validation Accuracy: 98.07%

Cell Types

The model can classify the following blood cell types:

1. Basophil - A type of white blood cell involved in inflammatory reactions
2. Eosinophil - White blood cells that fight parasites and allergic reactions
3. Erythroblast - Immature red blood cells
4. IG (Immature Granulocyte) - Immature white blood cells
5. Lymphocyte - White blood cells that fight infections
6. Monocyte - Large white blood cells that become macrophages
7. Neutrophil - Most common white blood cells that fight bacterial infections
8. Platelet - Cell fragments that help blood clotting

Training Details

• Training Images: 13,674
• Validation Images: 3,418
• Architecture: Pretrained ResNet-18 backbone with custom head
• Training Strategy:
  • 4 epochs with frozen backbone
  • 6 epochs with fine-tuning
• Input Size: 224x224 pixels
• Preprocessing: Standard ImageNet normalization

Performance

• Validation Accuracy: 98.07%
• All cell types: >95% precision and recall
• Best performers: Eosinophil and Platelet (100% precision)

Usage

from fastai.vision.all import *

# Load the model
learn = load_learner('cell_classifier.pkl')

# Predict on an image
pred, pred_idx, probs = learn.predict('path/to/blood_cell_image.jpg')
print(f"Predicted: {pred}")
print(f"Confidence: {probs[pred_idx]:.2%}")

Requirements

fastai>=2.7.0,<2.8.0
numpy<2.0
pillow>=10.0.0

Model Files

• cell_classifier.pkl - Complete fastai learner with model and preprocessing
• cell_classifier_weights.pth - PyTorch weights only
• confusion_matrix.png - Validation confusion matrix
• classification_report.csv - Detailed classification metrics
• training_summary.json - Training configuration and results

Citation

If you use this model, please cite:

@misc{pbc-cell-classifier-2024,
  title={ResNet-18 Peripheral Blood Cell Classifier},
  author={Your Name},
  year={2024},
  howpublished={Hugging Face Hub},
  url={https://huggingface.co/your-username/pbc-cell-classifier}
}

License

This model is released under the MIT License.

Created For

HuggingFace Agents-MCP-Hackathon Track 1 - MCP Tool/Server