tamara-kostova/multiagentmed-binary-tumor

# Brain Tumor Binary Classifier

PyTorch checkpoint artifacts for the MultiAgentMedClassifier binary brain tumor
MRI task. The repository contains a VGG16 CNN classifier checkpoint and,
optionally, a BiomedCLIP linear-probe checkpoint for classifying brain MRI
images as normal or tumor.

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

## Model Description

- Task: binary brain tumor MRI classification
- CNN architecture: VGG16
- Vision-language backbone for probe: `microsoft/BiomedCLIP-PubMedBERT_256-vit_base_patch16_224`
- Framework: PyTorch

## Classes

- `normal`
- `tumor`

The project-level BiomedCLIP labels are:

- `normal brain MRI`
- `brain tumor MRI`

## Files

- `binary_tumor/cnn/vgg16_MRI_tumor_binary_norm_final.pt`: VGG16 CNN checkpoint for binary brain tumor MRI classification.

• binary_tumor/biomedclip/linear_probe_BiomedCLIP_MRI_tumor_binary_norm_best.pt: BiomedCLIP linear-probe checkpoint for binary brain tumor MRI classification.

  Dataset

  Trained/evaluated for the binary tumor task using brain MRI tumor/normal data. The local evaluation script supports the Br35H binary layout:

  • data/Br35H/yes: brain tumor MRI
  • data/Br35H/no: normal brain MRI

  Update this section if you publish a model trained on a different dataset split or source.

  Training Details

  • Input size: 224 x 224 RGB
  • Normalization: ImageNet mean/std
  • CNN checkpoint: VGG16 fine-tuned for the binary_tumor task
  • BiomedCLIP probe: linear/MLP probe over frozen BiomedCLIP image features

  Metrics

  Evaluation is intended for the binary_tumor task on brain MRI tumor/normal datasets such as the Br35H binary layout described above. Recompute metrics on your held-out test set before using this model in a new domain or workflow.

  Inference Example

  Download the checkpoint from Hugging Face and point the local project config at it:

  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-binary-tumor",
      filename="binary_tumor/cnn/vgg16_MRI_tumor_binary_norm_final.pt",
  )
  
  DEFAULT_CONFIG.model.cnn_checkpoints["binary_tumor"] = checkpoint_path
  
  classifier = CNNClassifier(DEFAULT_CONFIG.model, DEFAULT_CONFIG.preprocess)
  result = classifier.classify("path/to/brain_mri.png", task="binary_tumor")
  print(result)
  

  For the BiomedCLIP probe:

  from huggingface_hub import hf_hub_download
  
  from agents.biomedclip_tool import BiomedCLIPTool
  from config import DEFAULT_CONFIG
  
  probe_path = hf_hub_download(
      repo_id="tamara-kostova/multiagentmed-binary-tumor",
      filename=(
          "binary_tumor/biomedclip/"
          "linear_probe_BiomedCLIP_MRI_tumor_binary_norm_best.pt"
      ),
  )
  
  DEFAULT_CONFIG.model.biomedclip_probe_checkpoints["binary_tumor"] = probe_path
  
  tool = BiomedCLIPTool(DEFAULT_CONFIG.model, DEFAULT_CONFIG.preprocess)
  result = tool.classify("path/to/brain_mri.png", task="binary_tumor")
  print(result)
  

  Intended Use

  This model is intended for research and experimentation in automated neuroimaging pipelines. It may be useful for prototype triage, benchmarking, and comparison against other image classifiers.

  It is not a medical device and should not be used as the sole basis for diagnosis, treatment decisions, or patient management.

  Loading In This Repository

  Use these files with this repository's local loaders:

  • CNN: config.ModelConfig.cnn_checkpoints["binary_tumor"]
  • BiomedCLIP probe: config.ModelConfig.biomedclip_probe_checkpoints["binary_tumor"]