Supportive Plane Correction (SPC)

• Supportive Plane Correction (SPC) is a set of methods for automatically aligning a reference plane—defined by parameters such as orientation, position, and surface normal—with various coordinate frames in medical imaging, instrumentation, and planetary science.
• SPC employs techniques including deep learning with 3D CNNs, feed-forward corrections, and iterative optimization to regress plane parameters accurately, utilizing representations like Euler angles, quaternions, and 6D matrices.
• The methodology has proven effective in reducing manual adjustments and workflow latency in intraoperative CT imaging, instrument tilt calibration, and planetary topography, with measurable improvements in precision and speed.

Supportive Plane Correction (SPC) comprises a set of methodologies for automatically or actively correcting the parameters of a reference plane—such as orientation, position, or surface normal—to bring it into functional alignment with either an anatomical, imaging, or instrumental coordinate frame. The term is utilized in distinct subdomains including intraoperative multiplanar reconstruction (MPR) alignment in CT imaging, synchrotron or scanning instruments’ plane calibration, and planetary topography (“stereophotoclinometry”), but is unified by its core objective: to estimate or control plane parameters so as to reduce manual adjustment, workflow latency, or measurement error.

1. Mathematical Formulation of Plane Representation and Parameterization

Across all applications, the mathematical representation of the plane is fundamental to SPC. In 3D medical imaging, each MPR plane is described by a point $\mathbf{A}\in\mathbb{R}^3$ (the in-volume center) and a pair of in-plane orthogonal direction vectors $\mathbf{e}_u,\mathbf{e}_v\in\mathbb{R}^3$. The plane normal is defined as $\mathbf{e}_w = \mathbf{e}_u \times \mathbf{e}_v$. Collectively, $$\mathbf{R}=[\mathbf{e}_u\;\mathbf{e}_v\;\mathbf{e}_w] \in SO(3)$$ provides the rotation matrix mapping the canonical plane basis to the global volume axes (Vicario et al., 2020); the plane equation is $\mathbf{n}^T \mathbf{x} + d = 0$, where $\mathbf{n} = \mathbf{e}_w$ and %%%%6%%%% (Vicario et al., 2021).

For rotation, several parameterizations are deployed:

• Euler angles (e.g., Z-X′-Z″ or z–y–x): To avoid discontinuities, SPC often regresses $$(\cos\alpha,\sin\alpha,\cos\beta,\sin\beta,\cos\gamma,\sin\gamma)\in[-1,1]^6$$.
• Unit quaternions: Four unconstrained components, normalized post-prediction, represent rotation.
• 6D matrix (“two-column”): The network regresses two unconstrained 3D vectors $r_1, r_2$, which are orthonormalized to span the rotation’s first two columns, with the third recovered by cross product. This approach avoids gimbal lock and normalization constraints, resulting in improved learning behavior and error metrics (Vicario et al., 2020, Vicario et al., 2021).

In the context of instrument tilt correction or planetary SPC, the plane is parameterized by tilt angle $\tau$ and azimuth $\varphi$ with respect to a rotation axis (Sereno et al., 2016).

2. Core SPC Methodologies in Imaging, Instrumentation, and Planetary Topography

Medical Imaging (MPR-Plane Correction)

SPC in intraoperative CT imaging relies on a feed-forward neural network (PoseNet-style 3D CNN) to regress per-plane translation and rotation parameters directly from volumetric data. The optimal regression is achieved via a loss function comprising

• $L_{\text{rot}}$: $\ell_2$ difference of rotation parameters,
• $L_{\text{trans}}$: $\ell_2$ difference of predicted vs. ground-truth centers,
• $L_{\text{orth}}$: penalty for non-orthogonality among the three anatomical planes.

Training is performed using He initialization, stochastic gradient descent with momentum, spatial/intensity augmentations, and 5-fold patientwise cross-validation (Vicario et al., 2020, Vicario et al., 2021). No segmentation labels are required—only annotated plane placements.

Instrument Tilt Correction (Active Azimuthal SPC)

For surface scattering or diffraction experiments, SPC denotes an active feedback method to maintain the incident or measurement angle at a fixed value despite residual tilt $\tau$ between the surface plane and the rotation axis. The key mechanism is a feed-forward correction: $$\theta_{\text{motor}}(\varphi) = \theta_0 + \tau\cos(\varphi+\varphi_0)$$ which exactly cancels the systematic $\cos\varphi$ modulation of the effective incidence angle, yielding stability at the $10^{-3}$ degree level (Sereno et al., 2016).

Stereophotoclinometry (SPC) for Planetary Surfaces

Classical SPC in planetary science produces high-resolution topography and albedo by tiling an object’s surface with small overlapping planar digital terrain models (“maplets”). Each maplet tracks local slopes $(p,q)$ and a mean albedo $a$, and optimization alternates between bundle-adjustment of camera poses and photometric correction of maplet surface parameters based on a radiance-reflectance model (e.g., Lunar–Lambert) (Driver et al., 11 Apr 2025). SPC alternates nonlinearly between pose and maplet updates rather than performing a monolithic factor-graph solution.

3. Algorithmic Architecture and Loss Functions

For deep-learning-based SPC in CT imaging, a typical architecture consists of five 3D convolutional blocks (Conv3D–BatchNorm–ReLU–Pooling), followed by three fully connected layers. The output dimension is determined by the number of planes $N_p$ and the rotation parameterization: $$\text{Output dimension} = N_p \times (3+\text{dim}_{\text{rot}})$$ where $\text{dim}_{\text{rot}}$ is 6 (matrix), 4 (quaternion), or 6 (Euler sin/cos).

The total loss enforces correct geometry: $$L = \alpha L_{\text{rot}} + \beta L_{\text{trans}} + \gamma L_{\text{orth}}$$ with hyperparameters typically optimized per anatomical region. $L_{\text{orth}}$ penalizes deviation from mutual orthogonality among anatomical planes, critical for clinical review.

For multi-anatomy learning, parameter sharing with a multi-head architecture enables lower memory cost and reduced overfitting, with each anatomical region having its own small head atop a shared 3D-CNN trunk (Vicario et al., 2021).

In planetary terrain SPC, the cost function comprises spatial (pixelwise orthoimage correlation) and photometric (reflectance law fit) terms, optimized in alternating Gauss–Newton or Levenberg–Marquardt loops (Driver et al., 11 Apr 2025).

4. Quantitative Performance and Comparative Analysis

Deep-learning SPC for CT achieves, on held-out test sets:

• Ankle region: median angular error in normal $E_n \approx 5.0^\circ$, in-plane rotation $E_i \approx 6.0^\circ$, translation $d\approx 6.0$ mm, and score $P \approx 6.67$.
• Calcaneus region: $E_n \approx 7.4^\circ$, $E_i \approx 7.7^\circ$, $d\approx 10.4$ mm, $P \approx 8.54$.

These figures are comparable to reported manual inter-rater variability ($E_n=6.3^\circ, d=9.3$ mm) and superior in speed and labeling overhead to segmentation-based approaches (Vicario et al., 2020). Multi-head architectures reduce position error $d$ from $7.4$ mm (single-task) to $6.1$ mm (multi-head), with orientation error unchanged at $\sim7^\circ$ (Vicario et al., 2021).

In active tilt-correction, SPC reduces incidence angle variations from $\pm\tau$ (up to $\pm1^\circ$) to $<10^{-3}$ deg, three orders of magnitude improvement, with residual error dominated by mechanical play (Sereno et al., 2016).

Classical planetary SPC yields peak signal-to-noise ratio (PSNR) values in the low 30s dB (e.g., Cornelia: $33.09$ dB, Ahuna Mons: $36.41$ dB), with photometric angular errors of $5$–$6^\circ$ and albedo within $6\%$. However, alternate methods such as PhoMo attain PSNR near 40 dB and tighter integration of terrain geometry (Driver et al., 11 Apr 2025).

5. Implementation, Labeling, and Computational Advantages

SPC’s clinical imaging pipelines execute inference in $<0.05$ seconds per volume on a standard GPU. Training is annotation-efficient; only plane placements are stored, with no need for segmentations or large-scale manual annotation. In practice, manual MPR plane adjustment by radiologists takes $46$–$55$ seconds per case (not including regloving), but SPC automates this step in fractions of a second, supporting real-time interaction in the operating room (Vicario et al., 2020).

In planetary applications, the traditional SPC pipeline demands expert oversight due to local minima and the multi-step iterative pose/maplet refinement process. New global optimization approaches reduce manual requirements at some expense of algorithmic complexity (Driver et al., 11 Apr 2025). Instrumental tilt SPC is hardware-feasible, requiring only per-sample calibration of tilt and phase, followed by real-time deterministic correction logic (Sereno et al., 2016).

6. Limitations and Future Directions

Medical SPC is constrained by rotation parameterization coverage—extreme or rare patient orientations are underrepresented. Current anatomy-conditional schemes (simple one-hot injection) do not improve over parameter sharing; advanced conditioning (e.g., FiLM, attention) is proposed for future work. Orientation error remains at $7^\circ$; further gains may arise from end-to-end coupling of all anatomical planes, transformer-based alignment, uncertainty modeling, or domain adaptation (Vicario et al., 2021).

Planetary SPC faces limitations from inflexible local maplet adjustment, requiring high-quality initial shape models and persistent human supervision. Full graph-based or joint optimization strategies (PhoMo) offer improved accuracy, mitigating height drift and photometric error and facilitating autonomy in mission planning (Driver et al., 11 Apr 2025).

Instrumental SPC is limited by mechanical backlash, surface inhomogeneities (splitting of the specular spot), and the need for per-sample calibration of tilt and phase; feedback is purely feed-forward, not adaptive (Sereno et al., 2016).

7. Domain-Specific Variants and Emerging Methodologies

The core methodology of SPC is adapted per scientific context:

• Intraoperative CT: Matrix-6D rotation regression, 3D CNN trunk, multi-head regression for standard planes.
• Instrumental tilt: Feed-forward angular correction based on calibrated tilt amplitude and phase.
• Planetary remote sensing: Alternating least-squares for pose and maplet updates, leveraging both photometric and geometric constraints.

Emergent approaches integrate deep learning for dense correspondence, factor-graph optimization across all variables (landmarks, normals, albedo, pose), and joint bundle-adjustment. This suggests future SPC-like solutions will employ interdisciplinary pipelines, drawing from computer vision, robotics, and domain-specific radiometric physics.

---

Key references:

• "Automatic Plane Adjustment of Orthopedic Intraoperative Flat Panel Detector CT-Volumes" (Vicario et al., 2020)
• "Active correction of the tilt angle of the surface plane with respect to the rotation axis during azimuthal scan" (Sereno et al., 2016)
• "Stereophotoclinometry Revisited" (Driver et al., 11 Apr 2025)
• "Automatic Plane Adjustment of Orthopedic Intra-operative Flat Panel Detector CT-Volumes" (Vicario et al., 2021)