README.md

Configuration

Contains JSON configuration files organized in two types:

1. General Config (e.g., config_atlas.json and config_brats2d.json): Defines model, dataset, training parameters, optimizer, loss, and metrics. This must be pass as a parameter to the main.py.

{
  "name": "experiment_name",
  "n_gpu": 1,

  "model": {"name": "UNet3D", "params": {...}},

  "dataset": {
    "name": "ATLAS",
    "root_folder": "/path/to/data",
    "transforms": "config/atlas_transforms.json", <-- path of Transforms Config
    ...
  },
  
  "classes":{"0": name_class_0, "1": name_class_1, ...},
  
  "aggregated_regions": { <-- optional: define aggregated regions for metrics
    "region_name_1": [1, 2],  <-- e.g., combine classes 1 and 2
    "region_name_2": [1, 2, 3]  <-- e.g., combine classes 1, 2, and 3
  },

  "loss": {"name": "DiceLoss", "loss_kwargs": {...}},
  "optimizer": {"name": "Adam", ...},
  "scheduler": {"name": "StepLR", ...},
  
  "metrics": [
    {
      "key": "DSC", <-- identification name of the metric
      "name": "DiceMetric", <-- name of the TorchIO/custom metric class
      "params": { "include_background": false }, <-- params for the metric constructor
      "train": true, <-- if metric must be computed during training phase
      "val": true, <-- if metric must be computed during validation phase 
      "test": true <-- if metric must be computed during test phase
    },
    ...
  ]
}

2. Transforms Config (e.g., atlas_transforms.json and brats2d_transforms.json): Specifies preprocessing and augmentation pipelines using TorchIO transforms (to be passed in the general config under the key dataset.transforms).

{
  "preprocessing": [...],
  "augmentations": [...]
}

Best model selection (model_best.pth)

The best checkpoint is selected using the validation value of the first metric in the metrics list (if --validation flag is passed to the main.py).

• The metric being tracked can be either "<key>_mean" (mean across classes) or "<key>_aggregated_mean" (mean across aggregated regions), controlled by the --eval_metric_type parameter
• It is assumed to be maximized (higher is better)

For example, if the first entry has "key": "DSC", the trainer tracks either DSC_mean or DSC_aggregated_mean on the validation set, depending on the selected evaluation metric type.

---

QaTaCov + Text Embeddings Config

config_qatacov2d_textemb.json extends the standard configuration by adding dataset.text_emb_dir to point to a directory containing precomputed report embeddings:

dataset:
  name: QaTaCovTextEmb
  text_emb_dir: /path/to/Text_Embeddings/BioBERT

The dataset expects the embeddings directory to contain:

• reports_emb.npy: NumPy array of BioBERT embeddings
• reports_meta.json: metadata about the embeddings
• image_to_report_idx.json: mapping from image stem to report index

Use scripts/extract_textemb_biobert.py or the slurm job in jobs/ to generate these files.