HNGeoNatomyX for Head & Neck RT

• HNGeoNatomyX is an extraction pipeline that quantifies anatomical changes in head and neck radiotherapy using serial imaging, RT structures, and clinical variables.
• It employs rigorous preprocessing steps including rigid registration, automated contour segmentation, and computation of 43 geometric metrics across defined anatomical subregions.
• A fraction-specific machine learning workflow predicts replanning needs with high accuracy, achieving up to 82% accuracy and detecting 76% of true positives early.

Searching arXiv for the cited papers to ground the article in current research. HNGeoNatomyX denotes, in its most explicit documented usage, an extraction pipeline for quantifying anatomical changes in head and neck cancer radiotherapy and for predicting whether replanning will be needed from serial imaging, RT structures, and clinical variables (Rios-Ibacache et al., 18 Jul 2025). In adjacent technical summaries, the same label is also used for geometry-centric anatomy systems derived from HoloView and HUG-VAS. This suggests a broader association with anatomically structured geometric computation, but the clearest operational definition is the head-and-neck radiotherapy workflow built to describe anatomical change during treatment and to support fraction-specific machine-learning prediction of replanning need (Kaushik et al., 15 Jan 2025).

1. Clinical definition and pipeline architecture

HNGeoNatomyX was developed on a cohort of 150 head and neck cancer patients treated at the McGill University Health Centre. Its inputs are pre-treatment CT sim DICOM images and RT structures exported from Eclipse v.15, on-treatment CBCT DICOM and CBCT-RT structures acquired twice weekly, and clinical and demographic data including weight, toxicities, KPS, p16 status, TNM, and systemic therapy (Rios-Ibacache et al., 18 Jul 2025).

The preprocessing sequence is explicitly structured. Each CBCT is rigidly registered to the CT sim coordinate system in MIM MAESTRO. The thermoplastic mask on the CT sim is contoured semi-automatically with a custom 2D slice-by-slice Python routine. The CBCT Field-Of-View cylinder is reconstructed from fraction 1 using pydicom, and the CT sim body contour is cropped to that FOV. A subsequent Z-axis trimming step ensures that all CBCT body contours span the same superior-inferior range. Anatomical subregions are then identified automatically: the neck is defined between the inferior mandible slice and the most inferior non-shoulder slice minus a three-slice buffer, and the submandibular plane is defined per patient to capture 2D radii and chord lengths in that slice (Rios-Ibacache et al., 18 Jul 2025).

The metric-extraction core operates per fraction, with CT sim treated as fraction 0 and CBCTs as fractions 1 through 25. For each fraction, the system computes 43 geometrical metrics across six categories—Body, Mask, PTV, Mandible, Neck, and Submandibular—using Python 3.7 together with pyvista, scipy, and point-cloud-utils. The output is a per-fraction metric matrix for each patient. Downstream analyses then comprise univariate slope analysis, relative-variation analysis, and machine-learning models constructed at fractions 5, 10, and 15 (Rios-Ibacache et al., 18 Jul 2025).

2. Geometrical representation and metric formalism

The pipeline formalizes anatomy through surfaces and contour-derived measurements. Let $S_0$ be the pre-treatment body or structure surface and $S_t$ the corresponding contour at treatment fraction $t$. Let $A(S)$ denote the volume enclosed by surface $S$, with voxel size $\Delta x \Delta y \Delta z$. This framing permits metric families that combine volumetric, point-cloud, and slice-wise descriptions of anatomical change (Rios-Ibacache et al., 18 Jul 2025).

Volume metrics include body volume,

$$V_{\mathrm{Body}}(t)=\sum_{(i,j,k)\in body(t)} \Delta x \Delta y \Delta z,$$

with analogous voxel-counting definitions for neck volume $V_{\mathrm{Neck}}$ and the air volume between body and mask $V^{air}_{Body\rightarrow mask}$. A centroid displacement expression is also stated: $$C(S)=\frac{1}{A(S)}\int_S x\, dV,\qquad \Delta C(t)=\|C(S_t)-C(S_0)\|_2.$$

Point-cloud distances are central. The Chamfer distance is defined as

$S_t$0

and the Hausdorff distance as

$S_t$1

On each axial slice $S_t$2, directed Hausdorff distances $S_t$3 are aggregated into

$S_t$4

Target-related and rigid-reference distances expand the description beyond global body change. For PTV-to-body geometry,

$S_t$5

with an average counterpart $S_t$6, and with $S_t$7 and $S_t$8 denoting volumes of the PTV outside and inside the body surface, computed by implicit distance clipping. Mandible-to-body quantities $S_t$9, $t$0, $t$1, and $t$2 are defined analogously, using the mandible surface as rigid reference (Rios-Ibacache et al., 18 Jul 2025).

Neck and submandibular descriptors add shape-sensitive local structure. The neck is characterized by 3D radii $t$3, $t$4, and $t$5 relative to its centroid, by average 2D cross-sectional area $t$6, by surface area $t$7, and by compactness

$t$8

The submandibular region is described through chord lengths $t$9, $A(S)$0 and 2D radii $A(S)$1, $A(S)$2, $A(S)$3. Relative-variation features are finally defined as

$A(S)$4

or, if $A(S)$5, with $A(S)$6 used in the denominator (Rios-Ibacache et al., 18 Jul 2025).

3. Statistical analysis and machine-learning workflow

The univariate analysis treats each metric as a temporal signal. For each patient and metric $A(S)$7, a linear regression $A(S)$8 is fit up to fraction $A(S)$9, and the slope $S$0 is used as the rate of change. Mann–Whitney U tests are then performed at each fraction to compare slope distributions between replanned and non-replanned patients. Fractions with $S$1 are marked statistically different, and the corresponding AUC of the test statistic is reported with bootstrap confidence intervals. A complementary delta analysis uses histograms of $S$2 at each patient’s replanning fraction to quantify the magnitude of change observed in replanned cases (Rios-Ibacache et al., 18 Jul 2025).

The predictive pipeline merges all information available up to and including a chosen fraction. Inputs comprise raw metrics $S$3, relative variations $S$4, slopes $S$5, clinical variables, toxicity grades, and weight. Missing values are handled by forward-fill for ordinal toxicity, linear interpolation for geometrical metrics, and the mean at the time point for other missing variables. Categorical variables such as p16, TNM, subsite, and therapy type are one-hot encoded, cancer stage is ordinal encoded, and all continuous variables are Z-scored with StandardScaler (Rios-Ibacache et al., 18 Jul 2025).

Training and evaluation are fraction-specific. The data are split into 70% training-plus-validation and 30% held-out test. For a fraction $S$6, only patients whose replan fraction is at least $S$7 are included among replanned cases in training and validation, to simulate real-time prediction. Feature selection combines a filter stage, SelectKBest by p-value ranking, and a wrapper stage, Recursive Feature Elimination with a RandomForest base estimator. Two feature counts are tested: $S$8 and $S$9. Nine classifiers are screened—LR, KNN, DT, AdaBoost, NB, SVM, RF, ET, and XGB—using repeated stratified 5-fold cross-validation with 10 repeats, reporting average ROC AUC, accuracy, and 95% confidence intervals. Hyperparameters are tuned with GridSearchCV on the best-scoring algorithms (Rios-Ibacache et al., 18 Jul 2025).

4. Reported predictive performance

The best fraction-specific multivariate models are reported for fractions 5, 10, and 15. At fraction 5, the selected classifier is ExtraTrees with criterion $\Delta x \Delta y \Delta z$0 entropy and max_features $\Delta x \Delta y \Delta z$1 'sqrt', using 32 features selected by RF-RFE from a mix of raw values, $\Delta x \Delta y \Delta z$2 values, and slopes across Body, PTV, Neck, Mask, Mandible, and Submandibular metrics. The test set contains $\Delta x \Delta y \Delta z$3 patients, comprising 17 replanned and 23 non-replanned cases. The model achieves ROC AUC $\Delta x \Delta y \Delta z$4 and accuracy $\Delta x \Delta y \Delta z$5, and it flags 13 of 17 true positives, corresponding to an early-prediction sensitivity of $\Delta x \Delta y \Delta z$6 (Rios-Ibacache et al., 18 Jul 2025).

At fraction 10, the best model is ExtraTrees with criterion $\Delta x \Delta y \Delta z$7 gini, max_features $\Delta x \Delta y \Delta z$8 'sqrt', and $\Delta x \Delta y \Delta z$9, using 64 features from all categories. The test set contains $$V_{\mathrm{Body}}(t)=\sum_{(i,j,k)\in body(t)} \Delta x \Delta y \Delta z,$$0 patients, with 13 replanned and 23 non-replanned cases. The reported performance is ROC AUC $$V_{\mathrm{Body}}(t)=\sum_{(i,j,k)\in body(t)} \Delta x \Delta y \Delta z,$$1 and accuracy $$V_{\mathrm{Body}}(t)=\sum_{(i,j,k)\in body(t)} \Delta x \Delta y \Delta z,$$2. At fraction 15, the best model is an SVM with sigmoid kernel, $$V_{\mathrm{Body}}(t)=\sum_{(i,j,k)\in body(t)} \Delta x \Delta y \Delta z,$$3, and gamma $$V_{\mathrm{Body}}(t)=\sum_{(i,j,k)\in body(t)} \Delta x \Delta y \Delta z,$$4 auto, using 64 RF-RFE-selected features. The test set contains $$V_{\mathrm{Body}}(t)=\sum_{(i,j,k)\in body(t)} \Delta x \Delta y \Delta z,$$5 patients, with 9 replanned and 23 non-replanned cases, and the reported performance is ROC AUC $$V_{\mathrm{Body}}(t)=\sum_{(i,j,k)\in body(t)} \Delta x \Delta y \Delta z,$$6 and accuracy $$V_{\mathrm{Body}}(t)=\sum_{(i,j,k)\in body(t)} \Delta x \Delta y \Delta z,$$7 (Rios-Ibacache et al., 18 Jul 2025).

Fraction and model	Test set	Reported result
Fraction 5; ExtraTrees, criterion=entropy, max_features='sqrt'	$$V_{\mathrm{Body}}(t)=\sum_{(i,j,k)\in body(t)} \Delta x \Delta y \Delta z,$$8: 17 replanned, 23 non-replanned	ROC AUC $$V_{\mathrm{Body}}(t)=\sum_{(i,j,k)\in body(t)} \Delta x \Delta y \Delta z,$$9, accuracy $V_{\mathrm{Neck}}$0, 13/17 true positives flagged
Fraction 10; ExtraTrees, criterion=gini, max_features='sqrt', $V_{\mathrm{Neck}}$1	$V_{\mathrm{Neck}}$2: 13 replanned, 23 non-replanned	ROC AUC $V_{\mathrm{Neck}}$3, accuracy $V_{\mathrm{Neck}}$4
Fraction 15; SVM, sigmoid kernel, $V_{\mathrm{Neck}}$5, gamma=auto	$V_{\mathrm{Neck}}$6: 9 replanned, 23 non-replanned	ROC AUC $V_{\mathrm{Neck}}$7, accuracy $V_{\mathrm{Neck}}$8

Taken together, the best specific multivariate models for fractions 5, 10, and 15 yield testing scores of 0.82, 0.70, and 0.79, respectively, and early prediction identifies $V_{\mathrm{Neck}}$9 of the true positives. The paper concludes that the created metrics have the potential to characterize and distinguish which patients will necessitate RT replanning, and that they show promise in guiding clinicians to evaluate RT replanning for head and neck cancer patients and streamline workflows (Rios-Ibacache et al., 18 Jul 2025).

5. Clinical role, standardization prospects, and limitations

The intended clinical application is direct integration into the treatment-planning workflow. HNGeoNatomyX can be integrated into the TPS, for example Eclipse, to automatically import CBCTs, register them to CT sim, contour the mask, and compute geometry metrics after each fraction. A fraction-specific machine-learning model can then provide a replanning risk score in real time, guiding radiation oncologists and physicists on whether to schedule CT re-simulation, re-contouring, and dose re-optimization (Rios-Ibacache et al., 18 Jul 2025).

The standardization objective is explicit. Delta-percentage histograms and ROC-derived thresholds could form the basis of a standardized replan-trigger guideline, with the text giving examples such as $V^{air}_{Body\rightarrow mask}$0 or $V^{air}_{Body\rightarrow mask}$1 mm by fraction 10. This does not constitute a finalized protocol; rather, it indicates how the metric framework could be translated into operational thresholds once validated. A plausible implication is that the pipeline is designed not only for retrospective description of anatomical change, but also for prospective decision support (Rios-Ibacache et al., 18 Jul 2025).

The reported limitations are threefold. First, some non-replanned cases may have clinically needed a replan but were not triggered, producing false positives because of undocumented logistical factors such as mask availability or the number of fractions remaining. Second, the body-contour cropping step uses a global CBCT FOV estimate; for live per-fraction use, the pipeline should only consider contours up to the current fraction. Third, the sample size is limited for rare subsites, including salivary gland, nasal, and hypopharynx, so external validation on larger multicentre cohorts is required for generalizability. Proposed future work includes incorporating dosimetric recalculation on CBCT and radiomic features, extending the framework to adaptive replanning timing optimization rather than binary prediction alone, and conducting a prospective clinical trial to assess workflow efficiency gains, dose-delivery accuracy, and toxicity reduction (Rios-Ibacache et al., 18 Jul 2025).

6. Relation to adjacent anatomy-centric systems

The name HNGeoNatomyX also appears in summaries of technically distinct systems, and these contexts clarify the broader methodological landscape. In augmented-reality anatomy education, a HoloView-based platform is described as using a thin-client AR headset such as HoloLens 2 for head pose, eye gaze, and hand gestures, while a remote GPU-equipped server running NVIDIA OptiX on CUDA performs stereoscopic ray-casting and foveation. The architecture includes a distributed rendering pipeline, gaze-contingent Gaussian foveation with $V^{air}_{Body\rightarrow mask}$2, hybrid surface-volume rendering, and a gesture vocabulary including pinch, fist, hand-ray selection, dual-hand navigation, slicing-plane interaction, and Bioscope mode. Reported performance changes from raw stereo at $V^{air}_{Body\rightarrow mask}$3 and approximately 7 fps over Wi‑Fi to sustained approximately 60 fps with foveated rendering at one-third resolution per dimension, with round-trip latency below 50 ms and a user study reporting average knowledge gain of $V^{air}_{Body\rightarrow mask}$4, SUS $V^{air}_{Body\rightarrow mask}$5, and TLX $V^{air}_{Body\rightarrow mask}$6 (Kaushik et al., 15 Jan 2025).

In landmark-based anatomical segmentation, HybridGNet provides a different but closely related geometry-first paradigm. The 2021 formulation combines a CNN image encoder with a spectral-GCNN graph decoder to output $V^{air}_{Body\rightarrow mask}$7 anatomical landmarks for lungs, heart, and clavicles in chest X-rays, using a fixed adjacency that encodes contour connectivity and a Chebyshev spectral filter of order $V^{air}_{Body\rightarrow mask}$8. The reported result is accurate and anatomically plausible landmark-based segmentation that is more robust to image occlusions than standard alternatives. The 2022 extension adds Image-to-Graph Skip Connections, deep supervision, fixed graph unpooling, and broader evaluation under domain shift, pacemaker occlusions, tuberculosis-related pathology, and cardiothoracic ratio estimation. A central conclusion in that line of work is that high Dice scores from pixel-based models do not guarantee anatomical plausibility, because Hausdorff distance and visual inspection can still reveal broken contours, spurious islands, holes, or disconnected components (Gaggion et al., 2021, Gaggion et al., 2022).

A further geometry-centric strand is represented by HUG-VAS, a hierarchical NURBS-based generative model for vascular geometry synthesis and controllable editing. HUG-VAS parameterizes vessel surfaces as NURBS with $V^{air}_{Body\rightarrow mask}$9, decomposes shape generation into centerline diffusion and guided radial-profile diffusion, and supports zero-shot conditional generation by Deep Posterior Sampling from prompts such as point clouds, contours, and partial patches. It is trained on 21 patient-specific aortas and is reported to generate anatomically faithful aortas with supra-aortic branches, with biomarker distributions that closely match the original dataset and with directly watertight quadrilateral grids usable in OpenFOAM Navier–Stokes simulations (Du et al., 15 Jul 2025).

Across these adjacent systems, anatomically meaningful structure is encoded through explicit geometry—surface meshes, graph landmarks, NURBS control nets, or contour-derived metrics—rather than treated solely as unconstrained dense pixels. This suggests that HNGeoNatomyX belongs to a wider methodological shift toward representations in which anatomical validity, correspondence, and editability are built into the computational object itself.