Neuro Ai

• Class 0: Background
• Class 1: Ventricular system
• Class 2: Normal WMH (aging-related periventricular changes)
• Class 3: Abnormal WMH (pathologically significant lesions)

Model Architectures

⭐ Recommended: Pix2Pix + Attention Discriminator + Adaptive Hybrid Loss (V5) for best performance

Repository Structure

results_fold_2_var_5_bet_zscore/
└── models/standard_4class/fold_2
    └── best_dice_generator.h5      # 4-Class: Background, Ventricles, Normal, Abnormal

Quick Start

Installation

pip install huggingface_hub tensorflow numpy nibabel

Download Models

from huggingface_hub import hf_hub_download

# Download best performing model (V5)
model_path = hf_hub_download(
    repo_id="Bawil/neuro-ai",
    filename="results_fold_2_var_5_bet_zscore/models/standard_4class/fold_2/best_dice_generator.h5"
)

# Load model
from tensorflow.keras.models import load_model
model = load_model(model_path)

Inference Example

import numpy as np
from tensorflow.keras.models import load_model

# Load pre-trained model
model = load_model(model_path)

# Prepare input (256x256 grayscale FLAIR MRI, normalized)
# input_image shape: (batch_size, 256, 256, 1)
input_image = preprocess_flair(your_flair_image)

# Run inference
predictions = model.predict(input_image)

# Get class predictions
predicted_classes = np.argmax(predictions, axis=-1)
# 0: Background
# 1: Ventricles
# 2: Normal WMH (periventricular)
# 3: Abnormal WMH (pathological)

# Extract pathological lesions only
abnormal_mask = (predicted_classes == 2).astype(np.uint8)

Training Data

Dataset Composition

• Local Dataset: 300 MS patients (6,000 FLAIR MRI slices)

  • Demographics: 78 males, 222 females
  • Age range: 18-68 years
  • Scanner: 1.5-Tesla TOSHIBA Vantage

• Public Dataset: MSSEG2016 (15 patients, 750 FLAIR slices)

Annotations

• Expert annotations by board-certified neuroradiologists (20+ years experience)
• Four-class labeling: Background, Ventricles, Normal WMH, Abnormal WMH
• Approved by Ethics Committee (IR.TBZMED.REC.1402.902)

Data Split

• Training: 70% patients (local)
• Validation: 10% patients (local)
• Testing: 20% patients (local) + 40% patients (public)
• Strategy: Patient-level stratified split (no slice-level leakage)

Model Training

Configuration

• Framework: TensorFlow 2.11, Keras
• Optimizer: Adam (learning rate: 0.0002)
• Loss Functions:
  • Option 1: Weighted categorical cross-entropy
  • Option 2: Unified Focal Dice
• Epochs: 60 (with early stopping)
• Batch Size: 4
• Input Size: 256×256×1

Hardware

• GPU: NVIDIA RTX 3060 (12GB VRAM)
• Training Time: 3-4 hours per model (5-fold CV)
• Inference Time: ~35-40ms per image

Model Performance

Use Cases

Clinical Applications

• MS Lesion Quantification: Accurate measurement of disease burden
• Differential Diagnosis: Distinguish pathological from normal aging
• Longitudinal Monitoring: Track disease progression over time
• Treatment Response: Evaluate therapeutic efficacy
• Radiological Reporting: Reduce false positive alerts

Research Applications

• Baseline Comparisons: Standardized evaluation framework
• Method Development: Foundation for advanced segmentation approaches
• Multi-center Studies: Protocol for broader validation
• Reproducible Research: Complete implementation available

Limitations

• Single Modality: Trained on FLAIR MRI only
• Scanner Specificity: Primarily 1.5T TOSHIBA data
• Disease Focus: Optimized for MS patients
• 2D Segmentation: Slice-by-slice processing (no 3D context)
• Resolution: Fixed 256×256 input size

Model Card

Intended Use

• Primary: Automated WMH segmentation for research and clinical decision support
• Users: Radiologists, neurologists, researchers, AI developers
• Out-of-scope: Not FDA/CE approved; not for standalone clinical diagnosis

Ethical Considerations

• Privacy: All data anonymized per HIPAA/GDPR standards
• Bias: Limited scanner/protocol diversity may affect generalization
• Clinical Validation: Requires expert review before clinical use
• Transparency: Complete methodology and code openly available

Model Card Authors

Mahdi Bashiri Bawil, Mousa Shamsi, Abolhassan Shakeri Bavil

Citation

@article{bawil2026neuro,
  title={AI-Driven Multi-Parametric MS Lesion Analysis from T2-FLAIR Imaging: A Clinical Decision Support Framework for Neuroradiology},
  author={Bawil, Mahdi Bashiri and Shamsi, Mousa and Bavil, Abolhassan Shakeri},
  year={2026},
  note={Models: https://huggingface.co/Bawil/neuro-ai}
}

License

MIT License - See LICENSE

Additional Resources

• 📄 Paper: [Under Review]
• 💻 GitHub Repository: mri-ms-analyzer/neuro-ai
• 📧 Contact: [email protected] | [email protected]
• 🏥 Institution: Sahand University of Technology & Tabriz University of Medical Sciences

Acknowledgments

• Golgasht Medical Imaging Center, Tabriz, Iran for providing clinical data
• Expert neuroradiologists for manual annotations
• Ethics Committee approval: IR.TBZMED.REC.1402.902

Keywords: Artificial intelligence (AI), multiple sclerosis, neuroradiology, clinical decision support, automated lesion analysis, deep learning, clinical AI