AnatomyCarve: Anatomy-Aware Imaging Operations

• AnatomyCarve is a family of techniques that employ anatomy-aware operations to selectively update, clip, and synthesize anatomical models.
• These methods integrate preoperative imaging with intraoperative guidance using refined segmentation, VR visualization, and generative synthesis.
• Applications across sinus surgery, cardiac imaging, and radiography demonstrate improved spatial accuracy and tailored clinical signal extraction.

Searching arXiv for "AnatomyCarve" and the specified related papers to ground the encyclopedia entry.
AnatomyCarve denotes a family of anatomy-aware operations in medical imaging and visualization that selectively remove, refine, expose, or update anatomical structures while preserving clinically relevant spatial relationships. Across the literature covered here, the term is used in several technically distinct but conceptually related senses: intraoperative carving of a preoperative model in functional endoscopic sinus surgery, segment-aware clipping of volumetric data in virtual reality, label-space refinement of cardiac anatomy, anatomy-prior-constrained segmentation, steerable extraction of anatomical instances, and compositional control of generated anatomy [2402.11840][2507.05572][2111.09650][2011.08769].

1. Conceptual scope

The available literature uses “AnatomyCarve” in more than one technical sense. In sinus surgery, it refers to updating a preoperative $3$D model by carving away resected regions and replacing them with geometry inferred from intraoperative endoscopy. In VR visualization, it denotes segment-aware clipping, where only selected tissue classes are removed inside user-placed clipping shapes. In cardiac CT and MR, it appears as a refinement paradigm that subdivides or constrains anatomical labels so that downstream inference respects clinically meaningful structure. In chest radiography and instance parsing, the same logic appears as anatomy-aware supervision and steerable extraction, respectively. In physicalization, related workflows carve volumetric anatomy into manufacturable sliceforms [2402.11840][2507.05572][2212.02014][2512.17263][2011.05689].

Taken together, these works suggest that AnatomyCarve is less a single algorithm than a recurring design motif: anatomy is not treated as undifferentiated geometry, but as a structured object whose parts, boundaries, inclusion relations, and spatial context constrain what may be removed, updated, or synthesized. That motif is explicit in label-to-label cardiac refinement, in containment-aware pathology segmentation, in polar-coordinate vessel-wall regression, in query-steered $9$-DoF instance retrieval, and in diffusion guidance based on tissue-wise geometric moments [2111.09650][2112.01137][2107.01748][2509.08015][2606.06509].

2. Intraoperative carving of preoperative geometry

In navigated functional endoscopic sinus surgery, the most explicit geometric use of AnatomyCarve is the “Endoscopic Chisel” method for updating a preoperative patient model during tissue removal [2402.11840]. The problem setting is ablative: surgeons remove bone and soft tissue, while navigation systems continue to rely on a CT-derived geometry that reflects undisturbed anatomy only. The proposed update rule is
$$
D(X){t+1} = \mathrm{Update}(D(X)_t, I{t+1}, T_{t+1}),
$$
where $D(X)_t$ is a truncated signed distance function defined in CT/model space.

The pipeline begins with CT segmentation and a preoperative tracked endoscopic sequence to build $D(X)0$. Endoscope intrinsics $K$ and distortion are calibrated, frames are undistorted, the optical tracker is registered to CT via segmented anatomical markers, and checkerboard-based hand-eye calibration yields camera extrinsics $T$. For each pose, a preoperative depth map is rendered from the CT-derived surface. Intraoperative depth is then estimated from tool-free endoscopic frames using a learning-based monocular depth estimator specialized for sinus endoscopy and fine-tuned on the preoperative sequence. Because monocular depth is scale ambiguous, the predicted depths are scaled and registered to the preoperative model by RANSAC-ICP using Open3D. Residual-based change detection uses
$$
r(u,v) = d{\mathrm{intra}}(u,v) - d_{\mathrm{preop}}(u,v),
$$
and
$$
M(u,v)=
\begin{cases}
1,& |r(u,v)| > \tau_d\
0,& \text{else}
\end{cases}
$$
with $\tau_d = 1.0$ mm. Fusion is then applied only where the binary mask indicates modification.

The TSDF update uses binary mask weighting:
$$
w_{t+1,i+1}=m_{i,t+1},
$$
$$
D_{t+1,i+1}=\frac{W_{t+1,i}D_{t+1,i}+w_{t+1,i+1}d_i}{W_{t+1,i}+w_{t+1,i+1}},
$$
$$
W_{t+1,i+1}=W_{t+1,i}+w_{t+1,i+1}.
$$
This preserves the preoperative model outside changed regions while preferentially integrating new depth along rays associated with tissue removal. Surface extraction is performed by Marching Cubes at the zero level set.

Quantitative evaluation was performed on an ex vivo specimen over five sequential perforations. The preoperative baseline error was $0.233 \pm 1.3$ mm. Without updating, the error in the modified region increased monotonically from $1.875 \pm 3.7$ mm at Bite 1 to $5.567 \pm 4.6$ mm at Bite 5. With updating, the corresponding errors were $2.432 \pm 2.3$, $3.397 \pm 3.7$, $3.172 \pm 3.6$, $3.538 \pm 2.8$, and $3.280 \pm 2.6$ mm. A depth ablation using intraoperative CT yielded the lowest errors, while omitting RANSAC-ICP registration substantially worsened performance, reaching $9.580 \pm 5.1$ mm at Bite 5. The reported residual on unmodified anatomy for evaluation registration was $1.832$ mm. The paper therefore attributes the dominant error source to monocular depth quality, not to the fusion logic itself [2402.11840].

3. Segment-aware exposure, steerable retrieval, and physical carving

In VR, AnatomyCarve is a segment-aware clipping technique for $3$D medical images rather than a surgical update operator [2507.05572]. It uses an intensity volume $V(x)$, an aligned segmentation map $S(x)$, and a set of clipping spheres ${s_i}$, each with a per-segment binary mask $M_i:L\rightarrow{0,1}$. A voxel is clipped if it lies inside at least one sphere whose mask marks that voxel’s segment as clippable:
$$
\mathrm{Clipped}(x)=\bigvee_{i=1}^{n}\left[\mathrm{Inside}(s_i,x)\wedge M_i(S(x))\right].
$$
The opacity function is then
$$
C(x)=
\begin{cases}
0,& \mathrm{Clipped}(x)=1\
\tau(V(x)),& \mathrm{Clipped}(x)=0.
\end{cases}
$$
This generalizes clipping from geometry space to semantic tissue space: skin, fat, or muscle may be suppressed, while ribs or spine remain visible as landmarks.

The GPU pipeline computes the clipped opacity volume, applies $3$D FXAA anti-aliasing, estimates normals with a $3$D Sobel operator, and performs DVR with Blinn–Phong shading. Optional Local Ambient Occlusion was disabled during interaction. The VR interface uses asymmetrical bimanual interaction: the dominant hand positions and scales clipping spheres, while the non-dominant hand selects visible segments and toggles them between clippable and unclippable. In a novice study with $n=17$, the average time per recreated view was $176 \pm 58$ s, the ratio of inserted spheres was $1.88 \pm 0.55$, the ratio of segment toggles was $0.98 \pm 0.62$, SUS was $77.6$, and NASA TLX averaged $32.7$. A new metric, MAE of the first encountered segment, correlated more strongly with user rankings than RMSE, with $R^2=0.93$ versus $0.70$. In an expert study with $n=8$ neurosurgeons and residents, SUS was $75.3$; reported strengths were improved understanding of spatial relationships, while reported weaknesses included insufficient resolution for small structures and missing $2$D imaging views [2507.05572].

A related but distinct carve operation appears in Med-Query, which maps directly to interactive extraction of specific anatomical instances from CT [2212.02014]. Here the key mechanism is not clipping spheres but DETR-style query embeddings bound to anatomical labels through a weighted adjacency matrix in Hungarian matching. Each instance is represented by a $9$-DoF box $(x,y,z,w,h,d,\alpha,\beta,\gamma)$, and only the requested queries need be executed at inference. On rib parsing, the method reported an identification rate of $97.0 \pm 4.2\%$ and an instance segmentation Dice score of $90.9 \pm 7.4\%$, with full-pipeline latency $2.591 \pm 0.977$ s and detection latency $0.065 \pm 0.008$ s. This extends the carve idea from semantic clipping to label-unique, steerable instance retrieval.

Physical carving is represented by the “Slice and Dice” workflow for anatomical edutainment [2011.05689]. An octree slices imported CT, MRI, or voxelized mesh data into assemblable sliceforms, an integer program determines feasible assembly order, and a packing algorithm lays out slices, labels, and instructions on printable pages. The paper reports per-model material cost as $< €1$. In a feasibility study, mean assembly time was $02{:}17 \pm 01{:}05$ for an aneurysm CT sliceform and $14{:}45 \pm 06{:}47$ for nested spheres; in an educational study, participants averaged $7.86$ correct, $0.57$ incorrect, and $1.43$ doubting answers when identifying structures on pre-assembled sliceforms. This is not clipping in the rendering sense, but it preserves the same anatomy-aware principle of selective decomposition.

4. Label refinement and anatomically constrained segmentation

One major AnatomyCarve usage is label-space refinement. In coronary CTA, a $6$-label whole-heart segmentation is refined to a $10$-label map by combining extrapolation and parcellation networks [2111.09650]. The initial label set consists of LV cavity, LV myocardium, RV cavity, LA, RA, and aorta. The refined set adds pulmonary artery and subdivides the LA into body, left pulmonary veins, right pulmonary veins, and left atrial appendage. U-Net 2 extrapolates PA from the $6$-label input without PA, U-Net 3 parcellates the LA mask, and U-Net 4 learns the final image-to-label mapping directly. On a $30$-case vendor-balanced test set, the median Dice scores for all original $6$ labels exceeded $95\%$ in the refined model; the LA body improved from approximately $91\%$ to approximately $97\%$, the RV cavity from approximately $92\%$ to approximately $96\%$, PA reached approximately $89\%$, and LPV, RPV, and LAA each exceeded $90\%$. Manual correction was required for only $80$ extrapolation cases and $50$ parcellation cases, compared with $260$ initial labels.

In cardiac MR pathology segmentation, AnatomyCarve appears as a soft inclusion prior rather than a label-parcellation operator [2011.08769]. The anatomy-prior-based U-net with attention encodes the containment hierarchy no-reflow $\subseteq$ infarction $\subseteq$ myocardium. The total objective is
$$
L_{\mathrm{total}} = L_{\mathrm{AWCE}} + \lambda L_{\mathrm{neigh}},
$$
with $\lambda = 10^{-2}$, where the neighborhood penalty encourages overlap and containment between child and parent probability maps. On EMIDEC validation slices, dense U-net with cross-entropy yielded Dice $0.955 \pm 0.009$ for LV, $0.871 \pm 0.045$ for myocardium, $0.622 \pm 0.080$ for infarction, and $0.246 \pm 0.102$ for no-reflow. Attention plus AWCE improved these to $0.971 \pm 0.014$, $0.926 \pm 0.029$, $0.769 \pm 0.082$, and $0.535 \pm 0.153$, while adding the neighborhood penalty yielded $0.970 \pm 0.007$, $0.916 \pm 0.029$, $0.747 \pm 0.082$, and $0.538 \pm 0.143$. The penalty therefore chiefly improved the most difficult label, no-reflow, while acting as a soft plausibility constraint rather than a hard topology guarantee.

A harder topological prior is used for carotid vessel-wall segmentation in black-blood MRI [2112.01137]. Instead of Cartesian semantic segmentation, the method transforms each slice into polar coordinates around the artery center and regresses a lumen radius $r_l(\theta)$ and a positive thickness $\Delta(\theta)$, defining the outer wall as
$$
r_o(\theta)=r_l(\theta)+\Delta(\theta), \qquad r_o(\theta)>r_l(\theta).
$$
This guarantees a ring-shaped vessel wall by construction. On the public challenge test set, the best model achieved median Dice $0.813$ for vessel wall and median Hausdorff distances $0.552$ mm and $0.776$ mm for lumen and outer wall, respectively. A conventional Cartesian U-Net baseline had cross-validation median Dice approximately $0.489$, with lumen HD approximately $1.07$ mm and wall HD approximately $1.252$ mm.

AnyCXR extends anatomy-aware segmentation to chest radiographs under partial and synthetic supervision [2512.17263]. Its Multi-stage Domain Randomization engine generated over $100{,}000$ synthetic radiographs from CT volumes, and Conditional Joint Annotation Regularization enforced anatomical consistency in a latent space for $54$ anatomical structures across PA, lateral, and oblique views. The total loss is
$$
\mathcal{L}{\mathrm{Total}}=\mathcal{L}{\mathrm{Seg}}+\lambda_{\mathrm{Dist}}\mathcal{L}{\mathrm{Dist}}+\lambda{\mathrm{Recon}}\mathcal{L}{\mathrm{Recon}},
$$
with $\lambda{\mathrm{Dist}}=4$ and $\lambda_{\mathrm{Recon}}=2$. Zero-shot on real CXRs, group-level DSCs were $0.926$ for PA heart/vessels, $0.891$ for PA lung, $0.894$ for PA rib, and $0.943$ for PA vertebrae; for lateral views they were $0.967$, $0.982$, $0.837$, and $0.795$, respectively. The resulting masks supported automated cardiothoracic ratio estimation with AUROC $0.93$, spine-curvature assessment with AUROC $0.87$, and disease classification whose mean AUROC improved from $80.26\%$ to $82.30\%$ when anatomical priors were added.

5. Compositional synthesis and the question of which anatomy to control

AnatomyCarve also appears in generative settings, where the goal is to manipulate anatomical factors rather than segment them. In DAA-GAN, cardiac MR synthesis is driven by disentangled anatomy arithmetic on spatial anatomy factors extracted by SDNet [2107.01748]. Selected factors from different subjects are swapped, aligned by center of mass, refined by a localized noise injection network $\mathcal{J}$, and re-entangled with imaging factors through a generator $\mathcal{G}$ using AdaIN. The total loss is
$$
\mathcal{L}{total} = \mathcal{L}{adv} + \mathcal{L}{path} + \lambda{1}(\mathcal{L}{cons} + \mathcal{L}{bg}),
$$
with $\lambda_{1}=10$. On balanced ACDC data, augmentation with generated samples improved classification accuracy from $88.3_{1.6}$ to $91.4^{*}_{1.4}$ and segmentation Dice from $84.9_{5.0}$ to $86.5^{**}_{4.7}$. On the underrepresented ARV class in M&M(s), accuracy improved from $82.7_{0.6}$ to $86.0^{*}_{0.8}$ and segmentation Dice from $83.5_{3.7}$ to $85.2^{**}_{4.1}$. FID was $17.2$ on ACDC and $24.8$ on M&M(s).

CardioComposer generalizes this idea to unconditional diffusion over multi-tissue $3$D segmentation maps guided at inference time by differentiable geometric moments [2509.08015]. User-specified ellipsoidal primitives provide target zeroth, first, and second moments for selected tissues, producing losses for size, position, and shape:
$$
L_{\mathrm{size}}^{(k)} = | m_0^{(k)} - m_0^{*(k)} |2^2,
$$
$$
L{\mathrm{pos}}^{(k)} = | c^{(k)} - c^{*(k)} |2^2,
$$
$$
L{\mathrm{shape}}^{(k)} = | \Sigma_n^{(k)} - \Sigma_n^{*(k)} |_F^2.
$$
The guidance is injected into reverse diffusion without retraining the model. The training set comprised $596$ CT segmentations with $11$ channels, resampled to isotropic $2$ mm voxels and cropped to $128^3$. The method supports independent control over size, shape, and position, composition across multiple tissues, and mesh extraction for downstream FEM simulation; a reported case study altered RV size while keeping LV fixed.

The question of which anatomy should be carved out and represented most carefully is addressed directly in a low-label benchmark for $5$-class cardiac pathology prediction on ACDC [2606.06509]. Using segmentation-derived descriptors from RV, myocardium, and LV, the study found that myocardium is the strongest single-structure source of signal and that concatenating RV, MYO, and LV performs best overall. On the ALL representation, multinomial logistic regression achieved balanced accuracy $0.85 \pm 0.089$, RBF-SVM $0.87 \pm 0.051$, and random forest $0.86 \pm 0.097$, whereas a small ResNet-18 on raw slices reached only $0.41 \pm 0.074$. A label-shuffle control dropped to $0.230 \pm 0.057$, near chance. The reported conclusion is that under limited labels, representation dominates model sophistication.

6. Applications, limitations, and interpretive issues

A common misconception is that AnatomyCarve names a single method. The literature instead spans intraoperative geometry updating, immersive visualization, segmentation refinement, generative control, low-label representation design, and even physical anatomical fabrication [2402.11840][2507.05572][2107.01748][2011.05689]. Another misconception is that anatomy-aware constraints automatically guarantee correct anatomy. Several papers explicitly use soft regularization rather than hard enforcement: the pathology neighborhood penalty is weak by design, AnyCXR relies on latent consistency rather than explicit topology, and diffusion guidance can still produce disconnected or duplicated structures under strong constraints [2011.08769][2512.17263][2509.08015].

The limitations are correspondingly modality-specific. In sinus surgery, accurate camera poses, CT registration, and monocular depth are prerequisites; pose errors, specularities, occlusions, non-rigid changes, and limited FOV can produce incorrect carving, and voxel resolution, truncation distance, runtime, and memory were not reported [2402.11840]. In VR clipping, output quality depends on segmentation granularity and GPU rendering trade-offs; local ambient occlusion was disabled during interaction, and experts reported insufficient precision for small structures and the absence of $2$D companion views [2507.05572]. In steerable CT parsing, the pipeline is not end-to-end and segmentation remains dependent on detection accuracy, even though query-based inference is efficient [2212.02014]. In carotid MRI, center misplacement and low-contrast outer-wall boundaries remain the primary error sources [2112.01137]. In generative settings, weight tuning, topology preservation, and dataset bias remain active constraints [2107.01748][2509.08015].

The application space is correspondingly broad. The sinus-surgery work explicitly points toward a continuously updated patient-specific model and a digital twin trajectory for FESS [2402.11840]. The VR system targets education and surgical planning, especially context-preserving exposure of deep anatomy [2507.05572]. Cardiac label refinement supports atrial fibrillation ablation planning and pulmonary artery valve localization [2111.09650]. Anatomy-aware chest X-ray segmentation supports cardiothoracic ratio estimation, spine-curvature assessment, and disease classification [2512.17263]. Steerable query-based extraction supports rib, spine, and abdominal-organ parsing in CT [2212.02014]. Physical sliceforms target anatomical edutainment and patient communication [2011.05689].

Taken together, these works define AnatomyCarve as a technically plural but coherent research direction: anatomy is used as the organizing prior that decides what may be clipped, what must remain for context, what can be subdivided into clinically meaningful parts, what must remain topologically plausible, and which structures carry the most informative signal for downstream analysis.