Organ-at-Risk (OAR) Segmentation

OAR segmentation demands exceptional boundary fidelity because radiation dose gradients are steepest at tissue interfaces. A 2mm error in the spinal cord contour can result in catastrophic myelopathy. Hausdorff Distance and surface Dice metrics are prioritized over volumetric overlap scores. Modern systems employ active contour losses and conditional random fields as post-processing steps to enforce smooth, anatomically plausible boundaries rather than relying solely on per-pixel classification.

Unlike binary tumor segmentation, OAR systems must delineate 15-30 distinct anatomical structures simultaneously in a single inference pass. Architectures like nnU-Net and TotalSegmentator demonstrate this capability, segmenting 100+ structures in whole-body CT scans. Key design considerations include:

• Hierarchical label taxonomies to resolve anatomical overlap
• Class imbalance handling via Tversky loss or focal loss
• Anatomical prior integration using shape constraints and spatial relationships between organs

Radiotherapy workflows involve multiple imaging modalities—CT for dose calculation, MRI for soft tissue contrast, and PET for metabolic activity. Robust OAR systems must either generalize across modalities or employ modality-agnostic architectures. Techniques include:

• Isotropic resampling to normalize voxel spacing across acquisitions
• Bias field correction (N4ITK) to standardize MRI intensity inhomogeneities
• Domain adaptation and test-time augmentation to bridge modality gaps without retraining

OAR segmentation systems must communicate confidence to radiation oncologists. Epistemic uncertainty (model ignorance) and aleatoric uncertainty (data noise from partial volume effects) are estimated via Monte Carlo dropout or deep ensembles. Outputs include per-voxel confidence maps that highlight ambiguous boundaries—such as the optic chiasm or brachial plexus—where manual review is mandatory before treatment planning proceeds.

Segmentation outputs must conform to the DICOM RT Structure Set standard for seamless integration into treatment planning systems like Varian Eclipse or Elekta Monaco. This requires:

• Conversion of binary masks to closed planar contours per axial slice
• Preservation of spatial registration metadata (Frame of Reference UID)
• Compliance with IHE-RO interoperability profiles
• Support for both DICOM Segmentation Objects and legacy RT Structure Set formats

Unlike tasks with objective ground truth, OAR segmentation reference standards exhibit significant inter-rater variability—measured by Cohen's Kappa or Dice between annotators. Leading systems model this ambiguity using simultaneous truth and performance level estimation (STAPLE) to fuse multiple expert contours into a probabilistic consensus map. Training on consensus labels rather than single-rater annotations improves generalization to real-world clinical variation.

The macroscopic extent of a malignant tumor visible on imaging or clinical examination. GTV represents the primary target for radiation dose delivery.

• Defined by the radiation oncologist using CT, MRI, or PET fusion
• Serves as the foundation for expanding to Clinical Target Volume (CTV) and Planning Target Volume (PTV)
• Accurate GTV contouring directly influences OAR sparing strategies
• Often segmented using contrast-enhanced imaging to distinguish tumor from surrounding healthy tissue

A specialized DICOM object that stores regions of interest, contours, and anatomical structures defined for radiotherapy planning.

• Encodes OAR and target volumes as planar contours or binary masks
• Links each structure to the reference CT or MR image series via spatial registration metadata
• Supports ROI interpretation through standardized naming conventions and color assignments
• Serves as the interoperability standard between treatment planning systems and delivery machines

A boundary-based metric measuring the maximum distance from any point in one set to the nearest point in the other.

• Quantifies the worst-case segmentation boundary error
• Critical for OAR evaluation where a single large deviation could place dose on healthy tissue
• 95th percentile Hausdorff Distance (HD95) is commonly used to exclude outlier sensitivity
• Complements overlap-based metrics like Dice Score for comprehensive boundary assessment

A statistical measure of spatial overlap between a predicted segmentation mask and the ground truth annotation.

• Calculated as 2 × |Intersection| / (|Prediction| + |Ground Truth|)
• Ranges from 0 (no overlap) to 1 (perfect agreement)
• The most widely reported metric for OAR segmentation accuracy
• Sensitive to volume size—small OARs like the optic chiasm require higher Dice thresholds for clinical acceptability

A self-configuring segmentation framework that automatically adapts preprocessing, network topology, and post-processing to any given medical dataset.

• Eliminates manual architecture tuning across different OAR segmentation tasks
• Automatically determines optimal patch size, batch size, and normalization based on dataset fingerprints
• Achieves state-of-the-art performance on multi-organ CT and MR segmentation benchmarks
• Widely used as the baseline for OAR auto-contouring challenges and clinical deployment

A pre-trained nnU-Net model capable of fully automatically segmenting over 100 anatomical structures in whole-body CT scans.

• Provides baseline OAR delineation for structures including lungs, heart, kidneys, liver, and spinal cord
• Trained on a diverse multi-institutional dataset for robust generalization
• Outputs can be imported directly into treatment planning systems as DICOM RT Structure Sets
• Serves as a starting point for manual refinement, reducing contouring time by up to 70%