tamara-kostova/multiagentmed-ms

# Multiple Sclerosis Binary Classifier

PyTorch checkpoint artifacts for the MultiAgentMedClassifier MS task.
Contains a ResNet101 CNN checkpoint and a BiomedCLIP linear-probe checkpoint
for classifying brain FLAIR MRI images as normal or multiple sclerosis.

These are checkpoint files for the accompanying project loaders, not standalone
Transformers models.

## Model Description

- Task: binary MS brain FLAIR MRI classification
- CNN architecture: ResNet101
- Vision-language backbone for probe: `microsoft/BiomedCLIP-PubMedBERT_256-vit_base_patch16_224`
- Framework: PyTorch

## Classes

- `normal`
- `ms`

The project-level BiomedCLIP labels are:

- `normal brain FLAIR MRI`
- `multiple sclerosis brain FLAIR MRI`

## Files

- `ms/cnn/resnet101_MRI_ms_norm_final.pt`: ResNet101 CNN checkpoint for binary MS brain FLAIR MRI classification.

• ms/biomedclip/linear_probe_BiomedCLIP_MRI_ms_norm_best.pt: BiomedCLIP linear-probe checkpoint for binary MS brain FLAIR MRI classification.

  Training Details

  • Input size: 224 x 224 RGB
  • Normalization: ImageNet mean/std
  • CNN checkpoint: ResNet101 fine-tuned for the ms task
  • BiomedCLIP probe: linear/MLP probe over frozen BiomedCLIP image features (layer 6)

  Metrics

  Model	Accuracy
  ResNet101 CNN	59.7%

  Note: MS classification from FLAIR MRI is a challenging task; the relatively lower accuracy reflects the difficulty of distinguishing subtle white matter lesion patterns. Recompute metrics on your own held-out test set.

  Inference Example

  from huggingface_hub import hf_hub_download
  from agents.cnn_tool import CNNClassifier
  from config import DEFAULT_CONFIG
  
  checkpoint_path = hf_hub_download(
      repo_id="tamara-kostova/multiagentmed-ms",
      filename="ms/cnn/resnet101_MRI_ms_norm_final.pt",
  )
  DEFAULT_CONFIG.model.cnn_checkpoints["ms"] = checkpoint_path
  classifier = CNNClassifier(DEFAULT_CONFIG.model, DEFAULT_CONFIG.preprocess)
  result = classifier.classify("path/to/brain_flair.png", task="ms")
  print(result)
  

  Intended Use

  Research and experimentation only. Not a medical device. Always validate on your own held-out test set before using in any pipeline.