Heart-surface imaging stops assuming a perfect torso model: OpDis-PS folds operator error into the likelihood and leads on pathological activation at 40 dB
Synopsis
Electrocardiographic imaging infers heart-surface potentials from body-surface potentials, but the torso conductor model used for inversion is only an estimate that respiration, posture and subcutaneous fat can shift away from the actual recording condition. OpDis-PS models this operator discrepancy as a signal-dependent linear term, corrects the nominal operator and folds the residual uncertainty into a whitened likelihood, while the diffusion prior stays fixed across all operator conditions. Under simulated lung-conductivity variation and an added subcutaneous-fat layer at 40 dB measurement noise, it reaches the lowest relative error and the highest spatial and temporal correlations on both physiological and held-out pathological activation.
Interpretation
Heart-surface potential reconstruction can now proceed under a torso model that differs from the one that produced the measurement, instead of treating model error as signal-independent noise. Earlier diffusion posterior samplers, including DDRM and DPS, assume the forward operator supplied at inference accurately represents the measurement process; in electrocardiographic imaging the nominal operator is only derived from an anatomical model and assumed tissue conductivities. At 40 dB, under both lung-conductivity and subcutaneous-fat discrepancy, OpDis-PS achieves the lowest relative error and the highest spatial and temporal correlation coefficients on the physiological and pathological test sets, with activation-time and localization errors at or near the lowest.
The approach lets the statistics of the discrepancy vary with the cardiac signal, then folds the predictable part into the operator. A joint-Gaussian approximation-error model conditioned on the nominal signal yields a linear correction to the nominal operator and a residual covariance, with the linear predictor and residual covariance estimated offline from training heart-surface potentials paired with sampled recording operators. At reconstruction time the residual discrepancy covariance is added to the measurement-noise covariance and the observation model is whitened before a spectral posterior sampler runs, while the unconditional diffusion prior stays frozen and is not retrained across operator conditions.
Against DDRM, which shares the same prior, nominal operator and sampler, OpDis-PS keeps a consistently lower relative error. That comparison isolates the gain to the discrepancy-aware likelihood rather than to a stronger generative prior or a different sampler. The evaluation uses openCARP simulations with 1700 training, 100 validation and 250 physiological test heart-torso pairs, plus 720 sequences from three pathologies at four severity configurations on 60 geometries each, none of which enters training or validation.
Perspective
The evaluation rests on openCARP-simulated electrophysiology and covers only two discrepancy sources, lung-conductivity variation and a subcutaneous-fat layer, with reconstruction targeting heart-template potentials on the outer surface of the left and right ventricles.
Real recordings may contain several sources of mismatch, including variations not represented in the operator distribution used to estimate the approximation-error model; in practice that distribution would need to be built from plausible variation in anatomy, tissue conductivities, electrode placement and acquisition conditions, and in the experiments it is narrower than the distribution generating the test operators, with the degree of misspecification not systematically varied. The discrepancy model is linear-Gaussian and treats residuals as independent across time although a recording shares one operator; its linear predictor explains less of the lung-conductivity discrepancy than of the subcutaneous-fat discrepancy. Improvements are also metric-dependent, and clinical relevance and uncertainty calibration still need assessment on real recordings.
