GEPAR3D: Geometric Prior 3D Analysis

• GEPAR3D is a framework that integrates advanced geometric priors with deep learning and parametric modeling for efficient 3D analysis.
• It enables instance-aware segmentation in dental CBCT, regional parcellation in cardiac imaging, and rapid digitization in triple-GEM detectors.
• The methodology achieves high accuracy and efficiency, demonstrated by metrics like a 95% DSC in tooth segmentation and significant speed-ups in simulation.

GEPAR3D refers to multiple distinct methodologies across several disciplines, each leveraging advanced geometric priors and parametric modeling for efficient, accurate analysis of three-dimensional biomedical or physical data. The term has been most prominently applied in three domains: (1) instance-aware 3D tooth segmentation in Cone-Beam Computed Tomography (CBCT); (2) volumetric regional parcellation of the right ventricle in cardiac echocardiography; and (3) fast parametric digitization for Monte Carlo simulation of triple-GEM particle detectors. While the acronym and implementations differ by field, all GEPAR3D variants incorporate geometry-driven approaches to overcome standard limitations in segmentation, partitioning, or digitization.

1. Geometry Prior-Assisted 3D Tooth Segmentation

GEPAR3D for CBCT tooth segmentation is a unified framework that directly integrates statistical shape modeling and 3D deep watershed energy basin representations within an end-to-end deep neural segmentation system. The pipeline operates as follows (Szczepański et al., 31 Jul 2025):

1. Preprocessing: Raw CBCT data is reoriented, resampled to 0.4 mm isotropic resolution, intensity-clipped, and normalized.
2. Coarse Localization: A lightweight 3D U-Net extracts a binary teeth mask and ROI.
3. Dual-Decoder Backbone: The main encoder-decoder structure simultaneously outputs (a) per-voxel semantic probabilities over 32 tooth classes; and (b) a continuous 3D energy map plus a 3-channel directional field for instance segmentation.
4. Instance-Guided Segmentation: Local maxima in the predicted energy map, refined by directional cues, seed a 3D watershed to separate individual teeth, followed by majority-voting semantic labeling.
5. Multi-Task Loss Fusion: Segmentation and instance detection are jointly optimized via a composite loss:

$$L_{\text{total}} = \Lambda_1 L_{EDT} + \Lambda_2 L_{\text{seg}} + \Lambda_3 L_{\text{dir}}$$

where $L_{EDT}$ measures regression on energy distance transform, $L_{\text{seg}}$ fuses Geometric Wasserstein Dice Loss (GeoWDL) and weighted cross-entropy, and $L_{\text{dir}}$ constrains directional field accuracy.

A key innovation is the integration of a statistical dentition shape model (SSM), encoded via a 33×33 penalty matrix $M_{l,l'}$ reflecting population-based geometric relationships, into the GeoWDL. This loss penalizes anatomically implausible segmentations without enforcing rigid adjacency (Szczepański et al., 31 Jul 2025). The energy basin approach, with explicit 3D distance and direction predictions, improves segmentation of roots and narrow apices—areas typically challenging for voxel-wise or purely semantic architectures. The method delivers a Dice Similarity Coefficient (DSC) of 95.0% and recall of 95.2% on external datasets, outperforming alternatives especially in apical regions.

2. Volumetric Parcellation of Cardiac Right Ventricle

The GEPAR3D framework is also employed in regional geometric and functional partitioning of the right ventricle (RV) in 3D echocardiography (Bernardino et al., 2020). Here, the emphasis is on anatomically coherent, reproducible segmentation of the apical, inlet, and outflow regions for quantitative analysis:

1. Geodesic Landmark Assignment: Three anatomical landmarks (apex, tricuspid annulus, pulmonary annulus) are identified on the RV endocardial surface mesh at end-diastole (ED).
2. Scalar Field Computation: For each surface vertex $p$, geodesic distances $d_a(p)$, $d_t(p)$, $d_p(p)$ to the respective landmark sets are computed using a fast marching or graph-tracing algorithm.
3. Tetrahedralization/Laplace Interpolation: The surface is volumetrically meshed with TetGen, and each distance field is harmonically extended into the blood pool by solving the Poisson--Dirichlet problem via P$_1$-FEM:

$L_{EDT}$0

where $L_{EDT}$1.

1. End-Systole Transport: Tracking provides a bijection between ED and ES surface meshes. Volumetric parcellations are recomputed at end-systole (ES) after Laplace interpolation from the mapped boundary.
2. Regional Assignment: The region label for each interior voxel is defined by minimum extended field:

$L_{EDT}$2

1. Volume/Ejection Fraction Calculation: Volumes are aggregated over tetrahedra assigned to each region; regional ejection fractions are:

$L_{EDT}$3

Validation demonstrates intra-observer error of 5.6%, inter-observer error of up to 23% (regionally), and parcellation sensitivity of 83–85% for local circumferential remodeling. The workflow is implemented using established libraries (Sfepy, TetGen) and is robust under segmentation variability (Bernardino et al., 2020).

3. Parametric Digitization for Triple-GEM Detectors

In high-energy physics, GEPAR3D denotes a fast digitization engine for simulating triple-GEM (Gas Electron Multiplier) detector responses (Farinelli et al., 2019). The aim is rapid reproduction of charge, timing, and spatial signals at a fidelity comparable to full Garfield++ simulations but at orders-of-magnitude increased efficiency.

The simulation replaces per-electron field tracing with a highly parametric, process-separated chain:

1. Primary Ionization: Charged particle track ionization is simulated with Poisson statistics; cluster rates $L_{EDT}$4 and per-cluster electron PDF are extracted from Heed.
2. Drift and Diffusion: Gaussian broadening of electron clouds is modeled in four gas gaps, with transverse and longitudinal diffusion coefficients derived from Magboltz fits:

$L_{EDT}$5

Parametric values: $L_{EDT}$6, $L_{EDT}$7.

1. Avalanche Gain: Multiplication in three GEM foils is modeled using exponential Townsend fits, effective transparencies, and aggregate Polya-distributed gain:

$L_{EDT}$8

with $L_{EDT}$9 per GEM.

1. Signal Induction: Multiplied charges are distributed on readout strips by integrating a Gaussian over the strip width. Arrival time is sampled, RC-shaped (50 ns), and noise added per time bin.
2. Validation and Tuning: Parameters (gains, diffusion, noise) are tuned using Garfield++/Magboltz for field and transport properties, and test-beam data for empirical scaling.
3. Performance: The engine achieves $L_{\text{seg}}$0--$L_{\text{seg}}$1 speed-up over Garfield++ (0.5 ms vs. 500 ms/event), maintaining agreement within 30% for charge yield, number of fired strips vs. track angle, spatial and timing resolutions.

GEPAR3D enables robust detector simulation for both offline studies and real-time applications, retaining accuracy for charge centroid and $L_{\text{seg}}$2TPC methods (Farinelli et al., 2019). The process utilizes lookup tables for all parametric models.

4. Key Methodological Features

Domain	Geometric Prior Mechanism	Core Model	Data/Output Type
Tooth Segmentation (Szczepański et al., 31 Jul 2025)	Statistical Shape Model + GeoWDL	Deep U-Net + Watershed	Per-voxel tooth instance/class
RV Parcellation (Bernardino et al., 2020)	Landmark-based geodesic Laplace fields	Tetrahedral FEM	Regional RV volumes, Ejection Fractions
GEM Digitization (Farinelli et al., 2019)	Parametric field-based transport/gain	Monte Carlo, Analytics	Waveforms, charge, simulated hits

All GEPAR3D instantiations share a commitment to process decomposition, explicit geometric modeling, and leveraging parametric/analytic descriptions (either as priors or as surrogates for computationally intensive sub-models), thus enabling robust, interpretable, and efficient computational workflows in their respective problem spaces.

5. Validation, Performance, and Clinical/Experimental Relevance

In CBCT tooth segmentation, the DSC and recall advantages (95.0% and 95.2%, respectively) are statistically significant and driven by improved root apex segmentation—critical for orthodontic root resorption assessment and clinical decision-making (Szczepański et al., 31 Jul 2025). In functional RV parcellation, segmentation is validated against observer variability and synthetic deformation, permitting regionalized remodelling studies necessary for heart failure and arrhythmogenic risk stratification (Bernardino et al., 2020). The simulation engine for GEMs preserves spatial and temporal fidelity, crucial for particle tracking efficiency in HEP experiments, and is validated by test-beam correspondence of charge centroid spatial resolution ($L_{\text{seg}}$3 at $L_{\text{seg}}$4) and $L_{\text{seg}}$5TPC timing ($L_{\text{seg}}$6) (Farinelli et al., 2019). These outcomes directly facilitate advanced research and practical deployment in biomedical and detector physics domains.

6. Distinctions in the Use of "GEPAR3D" Across Fields

While named identically, the GEPAR3D methodologies are independently developed and context-specific:

• In deep learning-driven instance segmentation (CBCT), GEPAR3D combines learned spatial instance-awareness (energy basins, directional fields) with population-based shape priors, producing unified multi-class and instance labeling via shared architecture and loss design (Szczepański et al., 31 Jul 2025).
• In biomedical simulation (RV), GEPAR3D encodes anatomical proximity using geometrically interpolated landmark distances, interpreted through harmonic extension and mesh-based partitioning (Bernardino et al., 2020).
• In detector digitization, GEPAR3D refers to compact, parameterized surrogates for detailed transport, diffusion, gain, and readout, replacing slow micro-simulations with fast, physically motivated statistical models (Farinelli et al., 2019).

A plausible implication is that the term GEPAR3D, while field-specific in modeling detail, signals an overarching design philosophy: the explicit integration of geometric priors/regimes into efficient, end-to-end computational frameworks for 3D analysis.