Methodology#
This page describes the scientific methods used by autoDS-PET for automated Deauville Score computation from PET/CT images.
Image Preprocessing#
SUV Normalization#
Raw PET images, acquired in units of activity concentration (Bq/mL), are converted to standardized uptake values normalized by body weight (SUVbw):
where:
\(C\) is the voxel activity concentration in Bq/mL,
\(W\) is the patient body weight in kg,
\(D\) is the effective injected dose in Bq.
When the DICOM DecayCorrection field indicates that no decay correction was
applied at acquisition (value "NONE" or empty), the injected dose is
corrected for radioactive decay between administration and acquisition time
using the radionuclide half-life. Otherwise, the reported total dose is used
directly.
Intra-patient Registration#
PET SUVbw images are rigidly registered to the corresponding CT images using the Elastix framework (SimpleElastix). Registration parameters:
Similarity metric: Advanced Mattes Mutual Information
Resolution levels: 3 (multi-resolution scheme)
Max iterations per level: 512
Interpolation: Linear (final resampling)
The output PET image is resampled onto the CT grid, sharing identical spatial geometry (voxel spacing, origin, orientation, and matrix size).
See autods_pet.imaging.registration.rigid_register_pet_to_ct().
Anatomical Segmentation#
CT images are segmented using TotalSegmentator to obtain a multilabel segmentation map providing anatomical structures including vertebrae, aorta, heart, liver, femora, and humeri.
Region of Interest Extraction#
Six ROIs are derived from the TotalSegmentator segmentation and, where applicable, further refined using image processing operations. All morphological operations (erosion, dilation) are performed in physical space (mm) using signed Euclidean distance transforms, making them robust to anisotropic voxel spacing.
Bone Marrow (Lumbar Vertebral Bodies)#
The L3, L4, and L5 vertebral masks are extracted from the multilabel segmentation and combined via logical union. This mask is intersected with a binary vertebral body segmentation to isolate the vertebral body compartment, excluding posterior elements. The resulting mask is eroded by 3 mm to reduce partial-volume effects at cortical boundaries.
Metric: 95th percentile of SUVbw within the eroded mask.
Mediastinal Blood Pool (Thoracic Aorta)#
The aorta mask is restricted to the axial extent of the T4-T8 vertebral slab through slice-wise gating. Cardiac spill-in is mitigated by dilating the heart segmentation by 6 mm and subtracting the dilated mask from the aorta. The resulting mask is eroded by 4 mm to obtain an intraluminal ROI. When multiple disconnected components remain, only the largest connected component is retained.
Metric: Voxelwise median SUVbw.
Liver#
The liver mask’s largest connected component is retained and holes are filled using binary flood-fill with full connectivity. The filled mask is eroded by 10 mm to avoid partial-volume effects.
Metric: Voxelwise median SUVbw.
Long Bones (Femoral and Humeral Diaphyses)#
For each of the four long bones (left/right femur, left/right humerus), the mask is cropped to its central 60% of axial extent to approximate the diaphysis, excluding the epiphyses. Each cropped mask is then eroded (5 mm for femora, 4 mm for humeri) to exclude cortical bone. The four diaphyseal masks are combined via logical union.
Metric: 95th percentile of SUVbw within the combined mask.
Brain (Cortical Grey Matter)#
The brain mask (TotalSegmentator label 90) is extracted from the multilabel
segmentation and its largest connected component is retained. When
grey_matter_only is enabled (the default), the mask is eroded by the
configured cortical_thickness_mm (default 5 mm) and the eroded
white-matter core is subtracted from the original to yield a cortical grey
matter shell.
Metric: Voxelwise median SUVbw.
The brain median is divided by the liver median to produce the Brain-to-Liver Ratio (BLR), a continuous metric (not a Deauville Score) reported alongside the standard DS columns.
Reference: Aide N et al. (2026). Brain-to-liver ratio in FDG-PET/CT for myeloma response assessment. Eur J Nucl Med Mol Imaging. doi:10.1007/s00259-026-07844-z
Focal Lesion#
Focal lesion masks are provided as binary segmentations (semi-automatic approach). Patients without a focal lesion mask are assigned DS 1 for the focal lesion component.
Metric: Maximum SUVbw within the lesion mask.
Deauville Score Assignment#
For each target ROI the Deauville Score is assigned by comparing target uptake against the MBP and liver reference values. Five component scores and one continuous metric are reported:
BM_DS – Bone marrow (vertebral bodies)
LB_DS – Long bones (femoral/humeral diaphyses)
FL_DS – Focal lesion
PM_DS – Paramedullary disease
EM_DS – Extramedullary disease
BLR – Brain-to-Liver Ratio (continuous value, not a Deauville Score)
Scoring criteria:
Score |
Criterion |
|---|---|
DS 1 |
No measurable uptake (focal lesion only; absence of lesion) |
DS 2 |
Target uptake ≤ MBP |
DS 3 |
MBP < target uptake ≤ liver |
DS 4 |
Liver < target uptake ≤ 2 × liver |
DS 5 |
Target uptake > 2 × liver |
This yields five component Deauville Scores per patient (BM_DS,
LB_DS, FL_DS, PM_DS, EM_DS) plus the continuous
BLR (Brain-to-Liver Ratio), each independently derived from the
automated pipeline. Per-patient scores are saved to
output_dir/{patient_id}_results/DeauvilleScores/deauville_scores.csv.
Implementation#
The pipeline is implemented in Python using SimpleITK-SimpleElastix for image
processing and registration, NumPy for array operations, and TotalSegmentator
for anatomical segmentation. The software is available as an installable
Python package (autods-pet).