AnatomyArchive: CT Segmentation Workflow Toolbox
- AnatomyArchive is a Python-based CT imaging toolbox that integrates segmentation with a standardized workflow for dataset curation, radiomics, and cinematic rendering.
- It streamlines CT scan selection and anatomy validation by automating target volume selection, arm detection, prosthesis exclusion, and group-aware segmentation mask management.
- Its modular architecture, built on TotalSegmentator integration, supports GPU-accelerated near photo-realistic 3D rendering and robust voxel-based radiomics for analytical research.
Searching arXiv for the cited AnatomyArchive paper and closely related anatomy-archive resources. AnatomyArchive is a Python-based CT image analysis toolbox built on top of TotalSegmentator and designed to extend automated segmentation into a broader workflow for dataset curation, anatomy-bounded volume standardization, mask management, body composition analysis, radiomics, statistics, and segmentation-integrated cinematic rendering. Its stated capabilities include automatic target volume selection and deselection according to user-configured anatomies for volumetric upper- and lower-bounds, a knowledge graph-based and time efficient tool for anatomy segmentation mask management and medical image database maintenance, automatic body volume cropping, automatic arm-detection and exclusion for more precise body composition analysis in both 2D and 3D formats, robust voxel-based radiomic feature extraction and visualization, and a Python-based GPU-accelerated nearly photo-realistic composite rendering workflow. The package is presented as a research and educational resource rather than a standalone clinical system (Xu et al., 18 Jul 2025).
1. Definition and scope
AnatomyArchive addresses a practical problem in CT-based research workflows: segmentation alone does not solve scan selection, anatomy coverage verification, standardized region-of-interest definition, mask organization, feature extraction, or visualization. The package is therefore positioned not as a new segmentation model but as a segmentation-centered workflow layer around TotalSegmentator. In the underlying description, it is used to ingest CT data, reorient images into a standard orientation, execute segmentation, determine whether required anatomy is present and complete, define anatomy-based upper and lower bounds, crop or standardize the body volume, detect prostheses, detect and remove arms for body composition analysis, extract radiomic features, visualize results, and generate outputs suitable for downstream machine learning development (Xu et al., 18 Jul 2025).
The package solves several specific bottlenecks. First, it automates inclusion and exclusion of retrospective CT studies according to imaged anatomy coverage, a task that would otherwise require manual review. Second, it replaces flat integer-based label-map handling with hierarchy-aware segmentation management. Third, it provides a robustness-oriented workflow for voxel-based radiomics, which the software description identifies as sensitive to extraction settings such as histogram bin width and kernel radius. Fourth, it adds segmentation-integrated cinematic rendering in Python, described as a previously missing capability in common CT research toolchains (Xu et al., 18 Jul 2025).
2. Software architecture
The software architecture is described as consisting of 10 components, each associated with a distinct part of the CT analysis pipeline (Xu et al., 18 Jul 2025).
| Component | Role |
|---|---|
| workflowConfig | Dictionary-based workflow control |
| genericImageIO | CT input/output and format conversion |
| segModel | Segmentation through TotalSegmentator |
| volStandardizer | Anatomy-bounded volume standardization |
| datasetManager | Dataset maintenance and tagging |
| segManager | Segmentation mask management and hierarchy handling |
| featureAnalyzer | Body composition and radiomics |
| simpleStats | Statistical tests and reliability analysis |
| simpleGeometry | Auxiliary geometric operations |
| dataVisualizer | 2D/3D visualization and rendering support |
The package architecture is modular, but the control hub is the current workflowConfig, which is described as a dictionary object with a predefined structure. It specifies task execution, segmentation targets, anatomy-based bounds, prosthesis detection, body composition analysis, coarse or fine segmentation settings, and feature extraction settings. The software description notes that workflowConfig is currently an umbrella term for dictionary-based configuration variables and that a future release will use a unified nested dictionary to control the full pipeline (Xu et al., 18 Jul 2025).
AnatomyArchive also abstracts differences between TotalSegmentator v1.5.7 and v2.8.0. The segModel module provides version-aware task configuration, class-map access, and segmentation execution while preserving auxiliary class maps unless they are explicitly removed. This design matters because the package uses auxiliary labels for tasks such as arm detection, where structures like the humerus may not reside in the primary map (Xu et al., 18 Jul 2025).
3. Segmentation-centered volume standardization
A central function of AnatomyArchive is the standardization of analyzed CT volume by anatomy-defined upper and lower bounds. Users define reference anatomies for an upper bound and a lower bound, and the package computes the corresponding -plane indices. Bones are recommended as bound-defining references because they are easier to detect and less deformable. If the relevant anatomy is not found, the returned value is ; if it is present but cropped, the returned value is . These outcomes can then be written into dataset tags for inclusion, exclusion, or warning logic (Xu et al., 18 Jul 2025).
Before this stage, images are reoriented into a common orientation using nibabel-based orientation transforms, with PLS+ used as the default target orientation because it is described as more convenient for generic human anatomy viewing. After segmentation, AnatomyArchive performs body cropping and standardized volume selection, checks whether the reference anatomy or body is cropped, and can optionally detect prostheses. Cropping detection uses maximum-intensity projections of segmentation masks into transverse, coronal, and sagittal planes and tests whether the projected mask touches image borders (Xu et al., 18 Jul 2025).
Arm detection is integrated into this standardization workflow because arm musculature can confound trunk body composition estimates when patients do not raise their arms. The package first identifies all body extremities, separates connected components, and then uses humerus, ulna, and radius masks to distinguish arms from legs. A related function also helps determine whether the body is cropped relative to the configured upper and lower bounds (Xu et al., 18 Jul 2025).
The segmentation engine is inherited from TotalSegmentator, which the software description characterizes as allowing segmentation of more than 100 organs and tissues. AnatomyArchive exposes version-dependent task settings while using a unified task abstraction. It supports fine segmentation at 1.5 mm and coarse segmentation at 3.0 mm and deliberately does not support 6.0 mm models because they may be too blurry for small lesions (Xu et al., 18 Jul 2025).
4. Segmentation mask management and knowledge-graph representation
AnatomyArchive treats conventional segmentation label maps as insufficient for large CT archives because flat integer labels do not encode anatomical hierarchy, synonymy, or many-to-many group membership. Its segManager module therefore introduces a hierarchical dictionary plus a directed multigraph representation implemented with NetworkX. This graph can encode parent-child anatomical relations, group memberships, left-right distinctions, and multiple memberships through group-name-tagged parallel edges, yielding a multigraph rather than a simple tree (Xu et al., 18 Jul 2025).
The package provides predefined anatomy groups such as bone, digestive accessory, intestine, muscle, endocrine, parenchyma, vasculature, and urinary, as well as broader system-level groups such as cardiovascular, musculoskeletal, gastrointestinal, and digestive. It also includes a synonym dictionary to normalize anatomy names; examples given include interpreting “hip” as pelvic bone and supporting “autochthon” as a synonym for spinal erectors. A further high-level categorization into bone, lung, and soft tissue supports automatic CT window selection during visualization (Xu et al., 18 Jul 2025).
Storage is also architecture-aware. Instead of storing every mask as a dense volume, AnatomyArchive can represent a segmentation mask by the 3D coordinates of its positive voxels and serialize this representation with MessagePack and sparse COO encoding. If a mask contains voxels, the coordinate representation is approximately $3J$ in size rather than for a dense mask array. Reconstruction requires the coordinate array and original image shape; restoration of original HU values can additionally use stored gray values at mask coordinates or the full original image array (Xu et al., 18 Jul 2025).
A plausible implication is that AnatomyArchive treats segmentation not only as image output but as database object. In that sense, it moves CT analysis toward archive logic: masks become retrievable anatomical entities with explicit hierarchy, synonym control, compact storage, and graph-encoded relations rather than only voxel labels.
5. Quantitative analysis, radiomics, and visualization
The featureAnalyzer module supports both body composition analysis and radiomic feature extraction. For body composition workflows, the package can evaluate either a 2D central plane or a 3D anatomy-bounded volume. The configuration allows selected objects, upper and lower reference anatomies, prosthesis exclusion, and HU-range enforcement. The software description gives default tissue ranges of HU for normal attenuation muscle, HU for fat, and HU for fatty or low-attenuation muscle. If muscle-range enforcement is enabled, muscles can be divided into categories, and intramuscular fat can contribute to inter-muscular adipose tissue characterization (Xu et al., 18 Jul 2025).
Radiomics is implemented through PyRadiomics with wrapper-level modifications. AnatomyArchive supports conventional radiomics and voxel-based radiomics, including multiple labels with label-specific optimized histogram bin widths. Optimal bin width is obtained with Doane’s algorithm through NumPy histogram binning and then rounded to preferred values such as . The package also introduces Standardization-based Subset Average Pooling for voxel-based radiomic features, described as a robustness-enhancing strategy that pools feature vectors across a subset of extraction conditions after standardization. The software description notes that this is conceptually analogous to average pooling in convolutional neural networks, but over extraction conditions rather than local neighborhoods (Xu et al., 18 Jul 2025).
Statistical analysis is integrated through simpleStats, which supports automatic test selection based on normality, paired or independent design, and variance assumptions. The described methods include regular t-tests, Welch’s t-test, paired t-test, Mann-Whitney U test, Wilcoxon signed-rank test, f-test variants, ICC, OCCC, DeLong’s test, Kolmogorov-Smirnov test, AUC confidence intervals, and sample-size estimation for linear mixed models (Xu et al., 18 Jul 2025).
Visualization spans both diagnostic and presentation tasks. AnatomyArchive supports windowed 2D displays, process-control overlays, voxel-based feature maps, and segmentation-integrated cinematic rendering. CT windowing is described by
0
The rendering subsystem provides whole-volume rendering, depth-unaware composite rendering, and depth-aware multi-volume rendering using VTK. Presets can be imported from 3D Slicer XML files, and the package emphasizes nearly photo-realistic segmentation-integrated composite cinematic rendering as one of its distinctive functions (Xu et al., 18 Jul 2025).
6. AnatomyArchive in the wider archive and atlas landscape
AnatomyArchive belongs to a broader class of systems that treat anatomy as structured, queryable data rather than only as raw imaging. The Scalable Brain Atlas provides unified web-based access to 20 atlas templates in six species, with plugins for coordinate transformations, anatomical connectivity, fiducial points, and 3D reconstructions (Bakker et al., 2013). The Active Atlas combines probabilistic 3D anatomical models and texture classifiers for 28 mouse brainstem structures and uses those models for automatic histological alignment and annotation (Chen et al., 2017). OntoAna models human anatomy as an OWL ontology covering cardiovascular, digestive, skeletal, and nervous systems (Vashisth et al., 2012), whereas ApiNATOMY represents multiscale physiological circuit maps through nodes, links, lyphs, chains, scaffolds, JSON-LD, RDF/OWL, and Neo4j integration (Kokash et al., 19 Sep 2025).
Large anatomy datasets complement these infrastructure layers. AbdomenAtlas provides 20,460 three-dimensional CT volumes with 673K masks across 25 abdominal, thoracic-adjacent, vascular, and skeletal structures and is linked to the BodyMaps challenge (Li et al., 2024). RadGenome-Anatomy projects volumetric chest CT anatomy into 25,692 paired PA and LL radiographic studies with 210 canonical anatomy classes and more than 10 million anatomy masks (Ye et al., 17 May 2026). HistoAtlas offers a morphology-centered archive of 6,745 TCGA H&E slides across 21 cancer types with 38 interpretable histomic features, cell maps, and compartment-resolved statistical associations (Bannier, 17 Mar 2026).
Within this landscape, a plausible implication is that AnatomyArchive occupies the workflow layer between segmentation engines and anatomical archives. It does not attempt to replace a primary atlas, ontology, or large reference dataset; instead, it standardizes CT volumes, organizes masks through anatomy-aware graph structures, and converts segmentation outputs into curated, analysis-ready resources that can be reused in retrospective studies, radiomics pipelines, and machine-learning development (Xu et al., 18 Jul 2025).
7. Uses, limitations, and significance
The software description presents AnatomyArchive as particularly suitable for retrospective and opportunistic CT studies, body composition research, radiomics, educational visualization, and the preparation of standardized datasets for machine learning. The paper explicitly states that it does not train machine learning models itself; rather, it prepares curated, standardized, feature-rich CT data that can assist development of modern machine learning models (Xu et al., 18 Jul 2025).
Its strengths follow from this position. It combines anatomy-aware inclusion and exclusion logic, graph-based mask management, compact sparse storage, upper and lower volume standardization, arm and prosthesis handling, body composition measurement, robust voxel-based radiomics, integrated statistics, and segmentation-aware rendering in a single Python toolbox. It also abstracts away differences between major TotalSegmentator versions, which reduces friction in long-term CT analysis workflows (Xu et al., 18 Jul 2025).
The package is also described with several limitations. It depends on a heavy Python software stack; the current workflowConfig is not yet a single unified global schema; direct MessagePack-based support for voxel-based radiomics has not yet been implemented; and the software is explicitly framed for research and educational purposes rather than clinical deployment (Xu et al., 18 Jul 2025). These constraints do not diminish its archival significance, but they clarify its role: AnatomyArchive is best understood as a segmentation-centered curation, analysis, and visualization framework for CT, not as a universal anatomy ontology, not as a primary atlas repository, and not as a regulatory-grade clinical platform.
