Transmission Zero Compensation

Updated 8 April 2026

Transmission zero compensation is a set of methodologies that substitute traditional transmission-based measurements with data- and physics-driven strategies to achieve accurate system reconstruction, such as μ-map estimation in SPECT.
It employs deep learning segmentation, hybrid physics-learning models, and Fourier series-based error cancellation to correct artifacts in imaging and mechanical drive systems.
In power electronics, active compensation eliminates nonminimum-phase behaviors by removing RHP zeros, thereby improving control bandwidth and system stability.

Transmission zero compensation refers to the class of methodologies designed to perform error or artifact mitigation in systems where the typical source of necessary information—most commonly a transmission-based measurement—is absent, unavailable, or intentionally omitted. These approaches arise in contexts such as tomographic imaging (notably SPECT without CT-based attenuation maps), high-precision drive systems (eliminating or compensating for kinematic transmission errors), and dynamic system control (removing destabilizing transmission zeros in control transfer functions). Across domains, transmission zero compensation replaces external measurements or compensates for structural zeros via data-driven, physics-based, or advanced feedback strategies to achieve system fidelity and performance that is statistically indistinguishable from conventional transmission-based methods, while reducing cost, complexity, and risk of error.

1. Transmission-Less Attenuation Compensation in SPECT

Conventional single-photon emission computed tomography (SPECT) systems require an external transmission scan, typically an X-ray CT, to map photon attenuation and enable accurate quantification. Transmission zero compensation in this setting leverages the physics of photon scattering: scattered photons recorded in lower energy “scatter windows” encode spatial information about the object’s linear attenuation coefficient μ(𝑟). Methods such as physics-informed initial estimation followed by deep learning-based segmentation exploit this principle, reconstructing attenuation maps by:

Reconstructing scatter window data via OSEM (Ordered Subsets Expectation Maximization) to yield a preliminary μ estimate.
Applying a segmentation CNN (typically U-Net architecture) to classify voxels into tissue-type regions using this estimate, optionally concatenated with photopeak reconstructions.
Assigning each segmented region a predefined, empirically derived μ value based on CT segmentation, forming a piecewise-constant or soft-weighted attenuation map.
Integrating this map into standard OSEM activity reconstruction with exponential path-weighted attenuation correction.

This approach has demonstrated root-mean-squared error (RMSE) <0.02 cm⁻¹ for estimated μ compared to ground-truth CT attenuation maps (Yu et al., 2021). Task-based area under the ROC curve (AUC) for perfusion defect detection is statistically non-inferior (AUC difference <0.05, p>0.05) relative to CT-based AC in large-scale simulation and clinical studies, while outperforming uniform or no AC (Yu et al., 2023, Yu et al., 2024).

2. Deep Learning and Hybrid Physics-Learning Approaches

Cutting-edge transmission zero compensation methods integrate physical modeling with convolutional neural architectures. Three key approaches are prominent:

CTLESS employs multi-channel input multi-decoder U-Nets (McEUN), accepting both scatter- and photopeak-window reconstructions as input channels, and produces segmentation probability volumes for K tissue classes (e.g., lung, bone, muscle, background) (Yu et al., 2024). The output is a voxel-wise weighted sum of assigned μ_k based on the segmentation probabilities.
SLAC utilizes a similar U-Net with attention-gated skip connections for scatter-activity fusion, segmentation, and assignment of μ values, trained on co-registered SPECT/CT data (Yu et al., 2023).
CNNs are regularized via class weighting, dropout, and attention mechanisms to address class imbalance and minimize overfitting, and are typically trained with Adam optimizer, Glorot initialization, and cross-validated class weights.

Physics-based inversion may precede segmentation, solving a Poisson likelihood maximization for μ via an iterative OSEM for the scatter projections (Yu et al., 2021).

3. Transmission Error Compensation in Harmonic Drive Systems

In mechanical systems such as harmonic drive transmissions, kinematic transmission errors—arising from geometric imperfections and torsional compliance—adversely affect output accuracy and dynamic behavior. Compensation approaches decompose the measured error into a periodic “pure” part (ε_p), modeled as a finite-order Fourier series of input angle, and a “flexible” dynamic part (ε_s), modeled via data-driven neural predictors (Wu, 2023).

Compensation schemes include:

Pure-part injection: Real-time evaluation of the truncated Fourier series for ε_p is subtracted from the position command, effectively canceling synchronous error components.
Nonlinear Model Predictive Control (NMPC): Embeds a one-step-ahead nonlinear NN predictor of ε_s to formulate and solve model predictive control for command trajectory updates over a rolling horizon.
Frequency-domain loop shaping: Sensitivity function shaping (using proportional, notch, and acceleration feedback filters) minimizes resonance and attenuates error at dominant harmonics, targeting frequencies at which “transmission zeros” induce poor disturbance rejection.

Hybrid approaches combine synchronous error cancellation with feedback loop shaping to balance frequency-domain suppression and computational feasibility.

4. Transmission Zero Elimination in Dynamic Systems and Power Electronics

In power electronics and wireless power transfer (WPT) receivers, right-half-plane (RHP) zeros in the system transfer function introduce nonminimum-phase behavior, degrading phase margin, dynamic response, and stability. In series–series compensated WPT systems, the use of diode-bridge rectifiers results in such RHP zeros in the small-signal control-to-output transfer function due to the fixed current-source nature of the secondary coil (Li et al., 2021).

Transmission zero compensation is achieved by:

Replacing the passive diode bridge with an actively controlled full-bridge rectifier, enabling direct regulation of DC-link current.
Altering the system’s small-signal transfer function to eliminate the RHP zero, yielding a strictly minimum-phase response and enabling high-bandwidth, robust feedback control.

Experimental results confirm 5× improvement in closed-loop bandwidth and stability margin, monotonic voltage response to reference changes, and suppression of inverse response previously associated with the RHP zero.

5. Evaluation Metrics and Performance Validation

Across modalities, transmission zero compensation methods are validated using task-specific and fidelity-centric quantitative metrics, including:

Metric	Imaging (SPECT)	Drive Systems	WPT Receivers
Task ROC AUC	Perfusion defect detection (CHO observer)	—	—
RMSE/SSIM	μ-map and activity vs. CT-derived reference	—	—
Kinematic error NRMSE	—	ε_s repeatability, predictor fit	—
Step/transient response	—	—	Settling time, overshoot
Closed-loop stability margin	—	—	Gain/phase margin

Transmission-less AC in SPECT achieves RMSE(μ) ∼0.01–0.018 cm⁻¹, SSIM ∼0.96, and defect-detection AUC statistically equivalent to CT-based methods (Yu et al., 2023, Yu et al., 2024). Harmonic drive pure-part injection compensates the dominant synchronous error with >98% NRMSE residual error fit, while NMPC and loop shaping can jointly address dynamic error but at significant computational overhead (Wu, 2023). Elimination of RHP zeros in WPT yields order-of-magnitude improvements in loop performance (Li et al., 2021).

6. Limitations, Open Challenges, and Generalization

Transmission zero compensation methodologies inherit specific limitations:

Piecewise-constant μ assignment does not capture fine-grained or pathological variability (e.g., heterogeneous lung pathology).
Deep learning segmentation is contingent on quality of ground truth; errors in CT segmentation propagate through the pipeline.
Most methods evaluate performance with model observers or in simulation; further clinical or field validation—including human observer studies and multi-site implementation—remains necessary.
List-mode or event-based data utilization and adaptive, patient-specific attenuation modeling represent open directions for further improvement (Yu et al., 2021, Yu et al., 2024).

In mechanical and electronic systems, computational complexity of NN predictors and NMPC remains a barrier to real-time deployment, motivating research into lightweight models or hardware acceleration (Wu, 2023).

7. Impact and Future Directions

Transmission zero compensation offers a practical pathway to cost reduction, reduced complexity, and mitigation of artifact or error sources in complex measurement and control systems. Imaging methods such as CTLESS, SLAC, and related approaches have the potential to enable SPECT-only attenuation compensation on legacy and mobile systems, substantially broadening access and utility. In mechanical and dynamic systems, advanced model-based and data-driven compensation schemes afford new levels of precision without hardware redesign or additional calibration procedures.

Future work is directed toward:

Integration of adaptive μ estimation and continuous-valued regression for individualized imaging (Yu et al., 2021).
Exploiting richer data modalities, such as SPECT list-mode or multi-energy projections.
Scaling NN-based compensation to real-time constraints in drive systems and implementing hybrid compensation schemes.
Broader, multi-center clinical and long-term deployment studies to establish robustness and generalization across populations and devices.

The ongoing development and successful validation of transmission zero compensation methods across domains demonstrate their essential role in advancing quantitative accuracy and system performance in the absence of transmission-based or otherwise inaccessible ground truth.