tamara-kostova/multiagentmed-stroke

# Ischemic Stroke Binary Classifier

PyTorch checkpoint artifacts for the MultiAgentMedClassifier stroke task.
Contains a DenseNet169 CNN checkpoint and a BiomedCLIP linear-probe checkpoint
for classifying brain CT images as normal or ischemic stroke.

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

## Model Description

- Task: binary stroke CT classification
- CNN architecture: DenseNet169
- Vision-language backbone for probe: `microsoft/BiomedCLIP-PubMedBERT_256-vit_base_patch16_224`
- Framework: PyTorch

## Classes

- `normal`
- `stroke`

The project-level BiomedCLIP labels are:

- `normal brain CT`
- `ischemic stroke brain CT`

## Files

- `stroke/cnn/densenet169_CT_stroke_binary_norm_final.pt`: DenseNet169 CNN checkpoint for binary stroke CT classification.

• stroke/biomedclip/linear_probe_BiomedCLIP_CT_stroke_binary_norm_best.pt: BiomedCLIP linear-probe checkpoint for binary stroke CT classification.

  Training Details

  • Input size: 224 x 224 RGB
  • Normalization: ImageNet mean/std
  • CNN checkpoint: DenseNet169 fine-tuned for the stroke task
  • BiomedCLIP probe: linear/MLP probe over frozen BiomedCLIP image features (layer 6)

  Metrics

  Model	Accuracy
  DenseNet169 CNN	97.7%

  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-stroke",
      filename="stroke/cnn/densenet169_CT_stroke_binary_norm_final.pt",
  )
  DEFAULT_CONFIG.model.cnn_checkpoints["stroke"] = checkpoint_path
  classifier = CNNClassifier(DEFAULT_CONFIG.model, DEFAULT_CONFIG.preprocess)
  result = classifier.classify("path/to/brain_ct.png", task="stroke")
  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.