focalclick_pacs_wwwl7_zoom09_s50w50

FocalClick checkpoint trained for one-click abdominal CT screenshot segmentation (PNG-domain) using mixed WW/WL robustness training and aggressive zoom augmentation.

What This Is

• Model family: FocalClick SegFormer-B0-S2 (focalclick backend in this repo).
• Task: one positive click -> binary mask for the clicked target on the current 2D slice.
• Domain: PNG-rendered CT slices (not raw DICOM/HU tensors at inference).
• Checkpoint file: best.pt (PyTorch checkpoint from supervised fine-tuning).

Training Snapshot

• Run dir:
  • focalclick/output/training/supervised_qc_wwwl7_s50w50_zoom09_initfresh_e10_v2
• Initialization:
  • model weights initialized from coarse s1p50_w0p50 checkpoint (no optimizer/scheduler carryover)
• Best validation score:
  • best_val_dice = 0.8385377488
  • best_epoch = 9
• Sample counts:
  • train: 1,334,487
  • val: 198,695

Preprocessing Assumptions (Important)

This model was trained with:

• channel transform:
  • channel_mode = linear
  • channel_strong_factor = 1.5
  • channel_weak_factor = 0.5
• WW/WL robustness set (mixed7):
  • original abdominal window + 6 additional presets
• zoom robustness settings:
  • zoom_min_factor = 0.9
  • zoom_max_factor = 4.0
  • translation_frac = 0.2
  • crop_target_margin_frac = 0.2

Inference helper scripts in this bundle default to channel_mode=auto and read channel settings from checkpoint metadata.

Files

• best.pt: best supervised checkpoint.
• metrics.json: training history and best-metric summary.
• manifest_mixed7_summary.json: mixed7 manifest composition summary.
• wwwl6_generation_summary.json: WW/WL variant generation summary.
• qc_rules.json: parsed QC exclusion rules used in data filtering.
• helpers/: updated inference helper scripts.

Helper Usage (Repo-Native)

Single-click example:

python -m focalclick.helpers.run_single_click_inference \
  --image /path/to/image.png \
  --checkpoint best.pt \
  --x 256 --y 256

Multi-click example:

python -m focalclick.helpers.run_multiclick_inference \
  --image /path/to/image.png \
  --checkpoint best.pt \
  --click 256,256,+ \
  --click 240,240,-

Notes / Constraints

• For research/development use.
• Not validated for clinical deployment.
• Performance can degrade if input rendering/preprocessing differs from the training pipeline.