Segmentation Mask

Masks encode anatomical labels as integer values at each spatial coordinate. Binary masks use 0 (background) and 1 (foreground) to isolate a single structure, such as a tumor. Multi-class masks assign unique integers to each anatomical region—e.g., 1 for liver, 2 for spleen, 3 for kidney—enabling simultaneous segmentation of multiple structures in a single volume. This encoding directly impacts loss function selection during training, with binary cross-entropy used for single-class tasks and categorical cross-entropy or Dice loss variants applied to multi-class problems.

A segmentation mask maintains exact spatial correspondence with its source image, sharing identical dimensions and origin coordinates. For a CT volume of size 512×512×300 voxels, the mask occupies the same grid. This one-to-one mapping is encoded in the DICOM header via shared Frame of Reference UID and Image Position (Patient) tags. Any resampling, cropping, or affine transformation applied to the source image must be identically applied to the mask to preserve alignment—a critical requirement for training supervised deep learning models.

Integer labels in a mask derive meaning from an associated ontology or coding scheme. Common standards include:

• SNOMED CT codes for anatomical structures
• RadLex terms for radiology-specific findings
• TNM staging labels for oncology masks Without explicit semantic mapping, a label value of '3' is ambiguous across datasets. Production-grade medical AI systems store label-to-ontology mappings in structured metadata, enabling interoperability across institutions and compliance with FHIR and DICOM Segmentation IOD standards.

Segmentation outputs exist on a spectrum of certainty. Hard masks assign a single discrete class to each voxel via argmax thresholding, producing the final clinical delineation. Probabilistic masks retain per-class probability values (e.g., 0.87 probability of 'tumor' at voxel i,j,k), preserving model uncertainty. Probabilistic masks are essential for:

• Uncertainty quantification and risk stratification
• Ensemble model agreement analysis
• Deformable registration where soft boundaries improve convergence Clinical workflows typically store hard masks, while research pipelines retain probabilities for downstream analysis.

Mask quality is quantified by comparing predicted segmentations against ground truth annotations. The Dice Similarity Coefficient (DSC) measures volumetric overlap, while Hausdorff Distance (HD) captures boundary accuracy by computing the maximum surface-to-surface distance. For clinical acceptance, DSC values above 0.85 are typically required for large organs, while small lesions demand stricter HD thresholds. Surface Dice at a tolerance (e.g., 2mm) has emerged as a clinically intuitive metric that correlates with radiologist-perceived quality better than voxel-wise measures alone.

In longitudinal studies and multi-modal fusion, masks from different time points or modalities (CT, MRI, PET) must be co-registered into a common coordinate space. Deformable registration warps masks to account for patient positioning changes, organ deformation, and tumor growth between scans. The resulting aligned masks enable:

• Tracking lesion volume changes over treatment cycles
• Fusing functional PET masks with anatomical CT masks
• Building 4D atlases for radiation therapy planning Misalignment errors propagate directly into downstream quantitative analysis, making registration accuracy a critical quality gate.