training_metadata.json

{
  "config": {
    "normal_dir": null,
    "output_dir": "/content/training_runs/medserra_chest_ai",
    "mask_dir": null,
    "covid_qu_ex_dir": "/content/training_runs/medserra_chest_ai/kaggle_assets/anasmohammedtahir__covidqu",
    "covid_kaggle_dataset": "anasmohammedtahir/covidqu",
    "download_covid_kaggle": true,
    "hf_token": "***redacted***",
    "hf_model_repo": "iraqigold/Chest-XRay-Autoencoder",
    "hf_dataset_repo": "iraqigold/X-Ray-Images",
    "hf_upload_repo": "iraqigold/Chest-XRay-Autoencoder",
    "upload_to_hf": true,
    "download_hf": true,
    "autoencoder_checkpoint": null,
    "base_checkpoint": null,
    "organ_profile": "chest_xray",
    "stage": "all",
    "seed": 42,
    "image_size": 256,
    "val_ratio": 0.12,
    "batch_size": 128,
    "num_workers": 0,
    "ae_epochs": 0,
    "seg_epochs": 10,
    "target_total_seg_epochs": 30,
    "ae_lr": 0.0001,
    "seg_lr": 0.0002,
    "weight_decay": 0.0001,
    "synthetic_positive_prob": 0.75,
    "hard_negative_prob": 0.2,
    "min_synthetic_lesions": 1,
    "max_synthetic_lesions": 9,
    "threshold_min": 0.005,
    "threshold_max": 0.5,
    "threshold_steps": 80,
    "save_every": 3,
    "no_imagenet": false,
    "resume_existing_ae": true,
    "resume_segmentation": true,
    "require_locked_radar_resume": true,
    "upload_every_epochs": 3,
    "upload_on_best": true
  },
  "train_count": 15926,
  "val_count": 2171,
  "best": {
    "best_threshold": 0.5,
    "best_dice": 0.8899215459823608,
    "best_iou": 0.8567548394203186,
    "epoch": 21
  },
  "notes": [
    "Use complete_chest_ai_best.pth for deployment when validation Dice is highest.",
    "The application architecture must stay identical: ae_radar + surgeon_refiner.",
    "COVID-QU-Ex support uses infection masks only; lung masks are intentionally ignored.",
    "All real and synthetic masks are trained as one binary disease_area target, without disease classification.",
    "To support another organ later, add an OrganProfile and ROI extractor, then reuse the same two-stage training flow."
  ]
}