README.md

metadata

license: apache-2.0
tags:
  - image-classification
  - medical
  - malaria
  - africa
  - tanzania
  - dinov2
  - blood-smear
  - imbalanced-classification
datasets:
  - harvard-dataverse-malaria-tanzania
metrics:
  - balanced_accuracy
  - mcc
  - recall
  - auc
base_model: facebook/dinov2-base

Malaria Detector — DINOv2 fine-tuned on Tanzania Blood Smears 🦟

Fine-tuned facebook/dinov2-base for binary classification of blood smear images: malaria infected vs healthy.

Trained on real patient data from 5 Tanzanian health centres, with rigorous handling of class imbalance.

📊 Performance (test set)

Metric	Value	Note
Accuracy	99.44%	Can be misleading with imbalance
Balanced Accuracy	99.25%	⭐ Imbalance-robust
MCC (Matthews)	0.988	⭐ Best single metric
F1 Macro	99.40%	Equitable across classes
Recall (Sensitivity)	100.00%	⭐ Critical clinical metric
Specificity	98.51%
AUC-ROC	100.00%
AUC-PR	100.00%	Imbalance-robust AUC

⚖️ Class imbalance handling

Dataset has 1.64x imbalance (Paludisme: 62% vs Sain: 38%).

Techniques applied :

1. ✅ Class weights in loss function
2. ✅ Stratified split on subtype (not just binary label)
3. ✅ Aggressive data augmentation
4. ✅ MCC-based model selection (instead of accuracy)
5. ✅ Multiple imbalance-robust metrics reported

🚀 Quick Start

from transformers import AutoImageProcessor, AutoModelForImageClassification
from PIL import Image
import torch

processor = AutoImageProcessor.from_pretrained('Sadou/malaria-detector-dinov2-tanzania')
model = AutoModelForImageClassification.from_pretrained('Sadou/malaria-detector-dinov2-tanzania')

img = Image.open('blood_smear.jpg').convert('RGB')
inputs = processor(img, return_tensors='pt')

with torch.no_grad():
    outputs = model(**inputs)

probs = torch.softmax(outputs.logits, dim=-1)[0]
pred = torch.argmax(probs).item()

labels = ['Healthy', 'Malaria']
print(f'{labels[pred]} (confidence: {probs[pred]:.1%})')

📂 Dataset

Tanzania Malaria Blood Smear Dataset

• Source: Harvard Dataverse
• DOI: 10.7910/DVN/O2WVWA
• Images: 3,544 real blood smears
• Source: 5 health centres in Tanga region, Tanzania
• Microscope: 4K SONY IMX334 sensor (40X-2500X)
• Staining: Giemsa reagent
• Reference: Lufyagila et al. 2024, Data in Brief

Class distribution

Class	Count
Thick Infected	1,139
Thick Uninfected	1,071
Thin Infected	1,064
Thin Uninfected	270

⚠️ Disclaimer

This is a research tool, NOT a certified medical device. Any real diagnosis must be confirmed by a qualified microscopist or physician. Clinical validation is required before any real-world deployment.

📚 Citation

@misc{malaria-dinov2-2026,
  author = {Sadou Barry},
  title = {Malaria Detector — DINOv2 fine-tuned on Tanzania Blood Smears with Class Imbalance Handling},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/Sadou/malaria-detector-dinov2-tanzania}
}