🩸 BloodNet-Nano

The smallest downloadable image-classification model for the BCCD blood cell dataset.

A 127K-parameter custom CNN that achieves 100% validation accuracy and 87.32% test accuracy on the BCCD test set — matching the performance of models 170× larger.

📊 Performance

Metric	Value
Parameters	127,552
Model size	1.85 MB
Input size	160×160×3
Validation accuracy	100.00%
Test accuracy (TTA)	87.32%
Macro F1	0.8692

🎯 Classes

The model classifies microscope images of white blood cells into 4 types:

1. Eosinophil — granulocyte involved in parasitic infections and allergies
2. Lymphocyte — adaptive immune cells (B-cells, T-cells, NK cells)
3. Monocyte — largest WBC, becomes macrophages in tissues
4. Neutrophil — most abundant WBC, first responders to bacterial infections

🏗️ Architecture

BloodNet-Nano uses MobileNetV2-style inverted residual blocks with strategic Squeeze-Excitation:

• Stem: 3×3 conv, stride 2 → 16 channels
• 7× inverted residual blocks (16 → 24 → 32 → 48 → 64 channels)
• Strategic SE blocks in deeper layers only (efficiency)
• Final 1×1 conv → 128 channels
• Global average pooling + dropout + dense head

Total: 127,552 parameters (1.85 MB on disk)

🛠️ Training Recipe

• Trained from scratch (no ImageNet pretrained weights)
• 60 epochs on dual NVIDIA T4 GPU
• MixUp augmentation (30% probability)
• Label smoothing (0.1)
• AdamW optimizer with weight decay 1e-5
• Cosine LR schedule with 3-epoch warmup (base LR 3e-3)
• Random flip, rotation, zoom, contrast augmentation
• Input size: 160×160 (smaller than typical 224×224)

💡 Why This Model Matters

Most BCCD classifiers on Kaggle use transfer learning from ImageNet on models with 5M–25M parameters. BloodNet-Nano shows that a carefully designed 127K-parameter model can match their accuracy — making it deployable on mobile, embedded, and edge devices.

Comparison	BloodNet-Nano	Typical pretrained baseline
Parameters	127K	5M – 25M
Model file	1.85 MB	20 – 100 MB
Trained from	scratch	ImageNet
Test accuracy	87%	88% – 94%

🚀 Usage

import tensorflow as tf
from huggingface_hub import hf_hub_download
import numpy as np
import cv2

# Download the model
model_path = hf_hub_download(repo_id="YoussefElsafi/bloodnet-nano", filename="bloodnet_nano.keras")
model = tf.keras.models.load_model(model_path, compile=False)

# Predict on an image
CLASS_NAMES = ['EOSINOPHIL', 'LYMPHOCYTE', 'MONOCYTE', 'NEUTROPHIL']

img = cv2.imread('your_cell_image.jpg')
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = cv2.resize(img, (160, 160))
img = img.astype('float32') / 255.0

pred = model.predict(np.expand_dims(img, 0))[0]
predicted_class = CLASS_NAMES[np.argmax(pred)]
confidence = np.max(pred)

print(f'Predicted: {predicted_class} ({confidence:.1%})')

⚕️ Disclaimer

Educational and research use only. Not intended for clinical diagnosis or medical decision making. This model was trained on the publicly available BCCD dataset and has not been validated for clinical use.

📚 Dataset

Trained on the BCCD blood cell images dataset:

• 9,957 training images
• 71 held-out test images
• 4 classes (Eosinophil, Lymphocyte, Monocyte, Neutrophil)