PlasmoSENet — Malaria Parasite Detection
A custom CNN designed from scratch for automated malaria parasite detection in Giemsa-stained thin blood smear microscopy images.
Results
| Metric | Value |
|---|---|
| Test Accuracy (with TTA) | 97.93% |
| Test Accuracy (without TTA) | 98.01% |
| Sensitivity | 97.19% |
| Specificity | 98.65% |
| Precision | 98.57% |
| F1 Score | 97.88% |
| Parameters | 2,107,094 (~8 MB) |
Trained from scratch on the NIH Malaria Cell Images Dataset (27,558 images). Achieves near-parity with MobileNetV2 fine-tuned baseline (97.97%) while using 38% fewer parameters and no pretrained weights.
Architecture
PlasmoSENet integrates five domain-specific design elements:
- Multi-scale stem — parallel 3x3/5x5/7x7 convolutions calibrated to Plasmodium erythrocytic stage dimensions
- SE channel attention — stain-aware feature weighting in every residual block
- Hybrid residual design — standard convolutions (stages 1-2) + depthwise separable inverted residuals (stages 3-4)
- Stochastic depth — linear drop path (0.0 to 0.1) across 9 blocks
- Kaiming initialization — stable from-scratch convergence without pretrained weights
Usage
import torch
from plasmosenet import PlasmoSENet
from torchvision import transforms
from PIL import Image
model = PlasmoSENet(num_classes=2)
model.load_state_dict(torch.load("model.pth", map_location="cpu"))
model.eval()
transform = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
image = Image.open("cell_image.png").convert("RGB")
input_tensor = transform(image).unsqueeze(0)
with torch.no_grad():
output = model(input_tensor)
prediction = output.argmax(dim=1).item()
classes = ["Parasitized", "Uninfected"]
print(f"Prediction: {classes[prediction]}")
Training
git clone https://github.com/Svetozar-Technologies/LocalMedScan-Models.git
cd LocalMedScan-Models
pip install -r requirements.txt
python scripts/train_plasmosenet.py --data-dir test_data/malaria/cell_images
Citation
If you use this model, please cite the repository:
@software{plasmosenet2026,
title={PlasmoSENet: A Multi-Scale Squeeze-and-Excitation Residual Network for Malaria Detection},
author={Svetozar Technologies},
year={2026},
url={https://github.com/Svetozar-Technologies/LocalMedScan-Models}
}
License
MIT