Skip to content

Systematic CT Cropping

Overview

Percentage-based DVH metrics (like V95%, V20Gy) and volume-sensitive radiomics features are heavily influenced by the total scan volume. Variations in the scan field-of-view (e.g., one patient scanned from L3 to mid-thigh, another from T12 to knees) introduce noise that can obscure biological signals.

Systematic CT Cropping standardizes this variable. It uses anatomical landmarks (vertebrae, joints, organs) automatically detected by TotalSegmentator to crop every CT scan and its corresponding segmentations to consistent physical boundaries.

The Problem

Consider two patients receiving the same absolute dose to the same absolute volume of the bladder: * Patient A (Long Scan): Total body volume in scan = 18,000 cm³. V20Gy = 500 / 18,000 = 2.8% * Patient B (Short Scan): Total body volume in scan = 15,000 cm³. V20Gy = 500 / 15,000 = 3.3%

Even though the treatment is identical, the metrics suggest a significant difference solely due to the scan length.

The Solution

By cropping both scans to the same anatomical landmarks (e.g., "Superior: Top of L1", "Inferior: 10cm below femoral heads"), we substantially reduce field-of-view-related variation in the denominator: * Patient A (Cropped): Volume ≈ 12,000 cm³. V20Gy = 4.2% * Patient B (Cropped): Volume ≈ 11,500 cm³. V20Gy = 4.3%

This makes cohort-level comparison of percentage metrics more interpretable and statistically defensible, although residual anatomical differences (body habitus, prior surgery, segmentation variability) still need to be considered.

Supported Anatomical Regions

The pipeline supports five predefined cropping templates based on stable bony landmarks. These templates represent one reasonable standardization scheme; users should adapt boundaries to their clinical protocol if needed.

1. Pelvis (Default)

  • Target: Prostate, Rectum, Bladder, Gyn
  • Superior Boundary: Superior aspect of L1 vertebra.
  • Inferior Boundary: 10 cm below the Femoral Heads (most inferior point).
  • Code: region="pelvis"

2. Thorax

  • Target: Lung, Esophagus, Heart, Breast
  • Superior Boundary: Superior aspect of C7 vertebra (or Lung apex + 2cm).
  • Inferior Boundary: Inferior aspect of L1 vertebra (or Diaphragm/Liver dome - 2cm).
  • Code: region="thorax"

3. Abdomen

  • Target: Liver, Pancreas, Kidney, Stomach
  • Superior Boundary: Superior aspect of T12/L1 vertebra (Diaphragm level).
  • Inferior Boundary: Inferior aspect of L5 vertebra (Pelvic inlet).
  • Code: region="abdomen"

4. Head & Neck

  • Target: H&N Cancer, Brain, Thyroid
  • Superior Boundary: Superior aspect of Skull/Brain + 2cm.
  • Inferior Boundary: Inferior aspect of C7 vertebra or Clavicles - 2cm.
  • Code: region="head_neck"

5. Brain

  • Target: GBM, Brain Mets
  • Superior Boundary: Superior aspect of Brain + 1cm.
  • Inferior Boundary: Inferior aspect of Brain - 1cm.
  • Code: region="brain"

Configuration

Enable and configure cropping in your config.yaml:

ct_cropping:
  enabled: true              # Master switch
  region: "pelvis"           # "pelvis", "thorax", "abdomen", "head_neck", "brain"

  # Margins extend the field of view beyond the strict landmark
  superior_margin_cm: 2.0    # e.g., 2cm above L1
  inferior_margin_cm: 10.0   # e.g., 10cm below femoral heads

  # Downstream integration
  use_cropped_for_dvh: true       # Calculate DVH on cropped volumes
  use_cropped_for_radiomics: true # Extract features from cropped volumes
  keep_original: true             # Keep original uncropped files

How It Works (Technical)

The cropping process is implemented in rtpipeline.anatomical_cropping and runs as a dedicated pipeline stage (crop_ct) after segmentation.

  1. Landmark Extraction: The system scans TotalSegmentator outputs for key reference masks (e.g., vertebrae_L1.nii.gz, femur_left.nii.gz).
  2. Boundary Calculation: It computes the physical Z-coordinates (in mm) for the superior and inferior limits based on the selected region and configured margins.
    • Fallback Logic: If primary landmarks (e.g., C7) are missing, it attempts to use secondary structures (e.g., Lung apex) and logs a warning.
  3. Cropping:
    • CT Image: The primary planning CT NIfTI is cropped to the calculated Z-range.
    • Masks: All segmentation masks (TotalSegmentator, Custom, etc.) are cropped to match the new CT geometry.
  4. RTSTRUCT Generation: A new DICOM RTSTRUCT file (RS_auto_cropped.dcm) is generated containing all cropped ROIs.
  5. Metadata: A cropping_metadata.json file is saved in the course directory, recording the exact boundaries used and the files modified.

Output Files

  • Cropped CT: NIFTI/ct_cropped.nii.gz
  • Cropped Masks: Segmentation_TotalSegmentator/*_cropped.nii.gz
  • Cropped RTSTRUCT: RS_auto_cropped.dcm
  • Metadata: cropping_metadata.json

Quality Control

Cropping is a destructive operation. The pipeline includes automatic QC to ensure no target structures are accidentally truncated.

  1. Audit Report: Run rtpipeline audit-cropping (or check the QC stage outputs) to see a summary of all cropped cases.
  2. Warnings: The logs will flag any case where the calculated boundaries fall outside the original scan extent (clamping).
  3. Visual Check: Use the "Visualize" stage to generate PNG snapshots of the cropped geometry.