src/BrainIAC/SequenceClassification/README.md

MR Sequence Classification

Overview

We present the MR sequence classification training and inference code for BrainIAC as a downstream task. The pipeline is trained and infered on T1/T2/FLAIR/T1CE brain MR, with balanced accuracy and AUC as evaluation metric.

Data Requirements

• Input: single Brain MR sequence
• Format: NIFTI (.nii.gz)
• Preprocessing: Bias field corrected, registered to standard space, skull stripped
• CSV Structure:

  pat_id,scandate,label
  subject001,20240101,0    # 0:T1w, 1:T2w, 2:FLAIR, 3:T1CE
  

refer to quickstart.ipynb to find how to preprocess data and generate csv file.

Setup

1. Configuration: change the config.yml file accordingly.

   # config.yml
   data:
     train_csv: "path/to/train.csv"
     val_csv: "path/to/val.csv"
     test_csv: "path/to/test.csv"
     root_dir: "../data/sample/processed"
     collate: 1  # single scan framework
    
   checkpoints: "./checkpoints/sequence_model.00"     # for inference/testing 
   
   train:
    finetune: 'yes'      # yes to finetune the entire model 
    freeze: 'no'         # yes to freeze the resnet backbone 
    weights: ./checkpoints/brainiac.ckpt  # path to brainiac weights
   

2. Training:

   python -m SequenceClassification.train_sequence
   

3. Inference:

   python -m SequenceClassification.infer_sequence