aiXiv:2502.001
Chengshuai Yang, Xin Yuan
physics.comp-img
cs.CV
Submitted February 2026
Computational imaging systems routinely underperform because the assumed forward model diverges from the true physics. Here we prove two results. First, the Finite Primitive Basis Theorem: every linear, shift-variant imaging forward model admits an epsilon-approximate representation as a typed directed acyclic graph over exactly 11 canonical primitives. Second, the Triad Decomposition: every reconstruction failure decomposes into three root causes -- information deficiency, carrier noise, and operator mismatch -- with mismatch dominant across all validated modalities. Across seven modalities spanning three carrier families (optical photons, X-ray photons, and nuclear spins), autonomous correction recovers +0.8 to +10.7 dB of mismatch-induced degradation without retraining the solver. Hardware validation on real instruments confirms mismatch dominance.
aiXiv:2502.002
Chengshuai Yang
cs.SE
cs.AI
Submitted February 2026
The SolveEverything abundance engine is implemented through a 10-gear framework that provides systematic infrastructure for addressing complex problems across computational imaging and beyond. The 10 gears are: (1) Targeting System (LIP-Arena), (2) Outcome Contracts, (3) Compute Escrow, (4) Action Networks, (5) Data Trusts, (6) Decision Logs, (7) Two-Source Rule, (8) Compute + Energy, (9) Fairness Targets, and (10) Literacy. Together, these establish maturity levels and an industrial stack for reproducible, transparent scientific and engineering practice. We demonstrate deployment across 168 imaging modalities.
aiXiv:2502.003
Physics World Model Team
cs.CV
physics.comp-img
Submitted February 2026
State-of-the-art EfficientSCI loses 20.58 dB when its assumed forward operator deviates from the physical truth by just eight parameters. This operator mismatch is the default condition of every deployed compressive imaging system, yet no existing benchmark quantifies it. We introduce InverseNet, the first cross-modality benchmark for operator mismatch in compressive imaging, spanning CASSI, CACTI, and single-pixel camera (SPC). InverseNet evaluates 12 reconstruction methods under a four-scenario protocol across 27 simulated scenes and 9 real hardware captures, totalling over 360 experiments. A mediocre algorithm with a correct forward model outperforms a state-of-the-art network with a wrong one.
aiXiv:2502.004
Physics World Model Team
cs.CV
physics.optics
Submitted February 2026
Forward model mismatch -- sub-pixel mask misalignment and dispersion drift between the coded aperture and detector -- is unavoidable in deployed CASSI systems, yet even moderate perturbations degrade state-of-the-art mask-guided transformers (MST) by over 16 dB. We present a self-supervised two-stage differentiable calibration pipeline that recovers 5-parameter mismatch from a single measurement and nominal mask alone, requiring no ground truth. Stage 1 performs a coarse hierarchical grid search scored by GPU-accelerated GAP-TV; Stage 2 applies gradient refinement through an unrolled differentiable forward operator using a Straight-Through Estimator (STE) for integer dispersion offsets. Evaluating five reconstruction methods across four scenarios on 10 KAIST scenes, the work uncovers a mask-sensitivity spectrum: mask-guided transformers suffer catastrophic degradation yet recover ~48% of the oracle gap after calibration.
aiXiv:2502.005
Chengshuai Yang
physics.med-ph
cs.AI
Submitted February 2026
We describe the clinical implementation of the CT QC Copilot, an automated decision-support system for computed tomography quality control embodying a "system computes, physicist decides" model. The Copilot reduced per-scanner QC analysis time from ~67 minutes (manual) to 4.2 minutes (automated + review) -- a 94% reduction making AAPM TG-233 trending-based QC practical for large clinical operations. Western Electric rule-based drift detection identified degradation 3-6 months before threshold exceedance in 4 of 30 simulated scanners. All nine ACR-aligned metrics showed agreement within 1.2 HU for CT number and 0.10 mm for geometric measurements.
aiXiv:2502.006
Chengshuai Yang
physics.med-ph
cs.SE
Submitted February 2026
We present the Physics World Model CT Quality-Control Platform, an open-source, reproducible software framework for automated CT quality assurance. Three architectural contributions: (i) CasePacks -- versioned, phantom-type-specific workflow descriptors that decouple QA logic from scanner hardware; (ii) a four-layer threshold system codifying institutional policies without forking code; and (iii) immutable CommissioningBundles with SHA-256 signing and full audit trails. Validates nine ACR-aligned metrics with zero inter-run variance and sub-second computation. Extensible to PET/CT and SPECT/CT via new CasePacks.
aiXiv:2502.007
Chengshuai Yang
cs.CV
physics.comp-img
math.FA
Submitted February 2026
Computational imaging forward models are traditionally implemented as monolithic, modality-specific codes. We prove that every forward model in a broad, precisely defined operator class admits an epsilon-approximate representation as a typed directed acyclic graph (DAG) whose nodes are drawn from a library of exactly 10 canonical primitives: Propagate, Modulate, Project, Encode, Convolve, Accumulate, Detect, Sample, Disperse, and Scatter. The proof is constructive and the library is minimal: removing any single primitive causes at least one modality to lose its representation. Empirical validation on 31 imaging modalities confirms that all achieve approximation error below 0.01 with at most 5 operator nodes and depth 5.