hongtai91/medsiglip-tb-classifier

MedSigLIP TB Classifier

Fine-tuned MLP classifier head on top of Google's MedSigLIP-448 vision encoder for tuberculosis detection from chest X-rays.

Performance

Metric	Test Set
AUC	0.9975
Sensitivity	96.2%
Specificity	98.5%
F1 Score	0.944
PPV	92.7%
NPV	99.3%

Architecture

• Encoder: MedSigLIP-448 (frozen, 1152-dim embeddings)
• Classifier: 3-layer MLP (1152 → 512 → 128 → 2) with ReLU, dropout=0.3, batch norm
• Training: 20 epochs, Adam (lr=1e-3), CrossEntropyLoss
• Training time: 7.6 seconds on RTX 4090

Training Data

• TBX11K: 8,812 images
• Shenzhen CXR: 662 images
• Montgomery County: 138 images
• Total: 9,342 chest X-rays (7,426 train / 957 val / 959 test)

3-Way Benchmark

Method	AUC	Sensitivity	Specificity
Random baseline	0.231	0.0%	100%
MedSigLIP zero-shot	0.837	77.2%	75.7%
Fine-tuned (this model)	0.9975	96.2%	98.5%

Usage

from tbscan.classifier import TBClassifier
from PIL import Image

classifier = TBClassifier(
    model_id="google/medsiglip-448",
    classifier_path="classifier.pt"
)
classifier.load()

image = Image.open("chest_xray.png").convert("RGB")
result = classifier.classify(image)
print(f"TB Probability: {result['prob_tb']:.1%}")

Part of TBScan

Built for the MedGemma Impact Challenge 2026. TBScan is an AI-powered TB screening system for low-resource clinics.

⚠️ Research and educational purposes only. Not validated for clinical use.