Chapter Four · failure evidence
What Image Reconstruction & Tomography got wrong, from 74 dissertations
The records evaluate tomographic and image reconstruction pipelines spanning optical, ultrasound, magnetic resonance, and X-ray computed tomography modalities. Across these systems, reconstructions frequently fail due to simplified physical models, undersampling artifacts, optimization instability, and deep learning distribution shifts. These records come from PhD theses at 18 institutions, 2021 to 2026. Each links to its thesis. They were extracted by language models reading the full text, so treat each as a lead to read, not a verdict.
Simplified forward models and ray approximations fail to capture wave diffraction and scattering physics
Ray-theoretic and geometric optics approximations struggle when neglecting diffraction, wave propagation limits, and medium contrast variations, resulting in divergence or severe spatial blurring. Unmodeled phase shifts, ballistic photon effects, and probe geometry constraints further degrade acoustic, optical, and impedance tomography reconstructions.
Tried and failed
iterative bent-ray ultrasound waveform tomography applied to high-contrast cortical bone imaging. Outcome: did not converge. Reason: diverged and formed wave-like artefacts despite ground truth or segmented average background initialisation
Ultrasonic bone characterisation: towards in vivo proximal-femur imaging · Imperial
Lost to a baseline
mIDT with multiplexing produced higher VMSE and lower reconstruction fidelity than conventional IDT under standard exposure conditions due to non-uniform TF weight distributions
Model and learning-based strategies for intensity diffraction tomography · OpenBU
Tried and failed
electrical impedance tomography reconstruction applied to sub-surface lesion shape classification. Reason: Inverse problem ill-posedness and low spatial resolution blurred all target geometries into circular high-conductivity regions
Wearable, implantable health sensors based on graphene and 2D materials · UT Austin
Tried and failed
direct coherent summation of multi-view images applied to sparse multistatic 3D image reconstruction. Reason: sensor positioning errors caused significant phase offsets, resulting in severe image defocusing
Multistatic-polarimetric SAR for sparsely sampled 3D imaging · Cranfield
Tried and failed
zero spatial frequency component inclusion in depth fitting applied to diffuse optical tomography depth reconstruction. Outcome: worse than baseline. Reason: unmodeled ballistic photons at zero frequency skewed single-exponent decay constants and increased depth error
Optical Imaging Of Tissue Physiology With Exogenous Contrast Agents · Penn
Tried and failed
forward-modeling with unelongated geometric basis functions applied to 3D plasma radiation tomography. Reason: the basis set lacked elongated structures needed to match anisotropic spatial radiation profiles
Tried and failed
ray-optics backprojection reconstruction applied to light-field microscopic volume imaging. Reason: geometric ray-optics approximations fail to model wave diffraction limits required for sub-cellular resolution
Fourier Light-Field Microscopy: Design, Optimization, and Applications · Georgia Tech
Tried and failed
multi-plane target calibration in narrow depth-of-field imaging applied to microscopic tomographic optical field reconstruction. Outcome: no signal. Reason: narrow depth of field caused severe blurring of out-of-focus calibration markers across offset planes
Inter-scale energy transfer in turbulent premixed combustion · Georgia Tech
Tried and failed
aperture-based multiplexing in miniaturized optical probes applied to endoscopic optical coherence tomography. Reason: optical element thickness relative to probe diameter caused severe diffractive degradation
Volumetric Optical Imaging of Tissue Microstructure for Grading of Dysplasia In Vivo · MIT
Considered and rejected
Considered and rejected: Rejected standard ray-theory tomography (bent ray tomography alone) due to severe resolution loss from neglecting diffraction
Shear-horizontal guided wave tomography · Imperial
Considered and rejected
Considered and rejected: Rejected pure Diffraction Tomography (Born/Rytov approximations alone) due to contrast/size limitations and phase-unwrapping instability
Shear-horizontal guided wave tomography · Imperial
Considered and rejected
Considered and rejected: Rejected direct ray-based travel-time tomography due to low spatial resolution (Fresnel zone ~2.2 cm vs ~1 mm for FWI).
Tried and failed
rectangular sensor array layout with short baseline applied to tomographic wave-based damage reconstruction. Reason: short inter-sensor distances produced tiny separation coefficients and local reconstruction artifacts around transmitters
Pipeline health monitoring using helical guided ultrasonic waves · UT Austin
Tried and failed
filtered back-projection with ramp filtering applied to plane-wave ultrasound image reconstruction. Outcome: worse than baseline. Reason: missing low-frequency components in transducer Gaussian pulses negated ramp filter benefits
Tried and failed
spatial-domain microlens array light-field imaging applied to 3D volumetric micro-structure reconstruction. Reason: Severe reconstruction artifacts and loss of resolution near the native focal plane due to phase singularities.
Fourier Light-Field Microscopy: Design, Optimization, and Applications · Georgia Tech
Deep learning reconstruction models fail under distribution shifts and unmodeled physical data mismatches
Supervised deep neural networks and generative diffusion models overfit or collapse when encountering out-of-distribution targets, environmental shifts, or discrepancies between synthetic and experimental noise. Omitting domain adaptation or standard physical preprocessing leads to severe reconstruction error and loss of structural detail.
Tried and failed
synthetic procedural noise generation for training data applied to microscopy image denoising and artifact removal. Outcome: did not generalise. Reason: procedurally generated corruptions failed to accurately replicate experimental noise characteristics, degrading reconstruction quality
Accelerating microscopy for nanoscale fabrication via deep-learning image reconstruction · UT Austin
Lost to a baseline
dAUTOMAP reconstruction was blurrier and produced higher SSIM loss than Filtered Back Projection (FBP) on noiseless/high-resolution 256x256 and 512x512 CT images.
Design and applications of cold-cathode X-ray imaging systems · MIT
Tried and failed
deep learning based image reconstruction applied to snapshot Fourier ptychography. Outcome: did not generalise. Reason: training data lacked artistic images, leading to significant reconstruction errors on out-of-distribution inputs
Sampling in Computational Cameras · DukeSpace
Tried and failed
deep learning reconstruction without domain adaptation applied to radio tomographic imaging across environments. Outcome: did not generalise. Reason: Severe distribution shift and model mismatch between source and target environments degraded reconstruction accuracy.
Radio Tomographic Imaging with Deep Learning · Georgia Tech
Tried and failed
supervised deep neural networks applied to out-of-distribution light-field functional microscopy reconstruction. Outcome: did not generalise. Reason: distribution shift from structural training data due to noise and scattering mismatches
Computational methods for 3D imaging of neural activity in light-field microscopy · Imperial
Tried and failed
pure deep learning estimators without physics adaptation applied to tomographic reconstruction under model mismatch. Outcome: did not generalise. Reason: models lacked robustness to mismatched forward projection models across varying noise levels
Radio Tomographic Imaging with Deep Learning · Georgia Tech
Tried and failed
unsupervised deep learning without dataset-specific tuning applied to cross-dataset dynamic MRI reconstruction. Outcome: did not generalise. Reason: lack of task-specific retraining caused severe domain shift, producing black images and massive reconstruction error
Generalizable low-latency accelerated dynamic MRI · Iowa State
Tried and failed
retrospectively undersampled training for accelerated imaging reconstruction applied to prospective magnetic resonance image reconstruction. Outcome: did not generalise. Reason: prospective acquisition contains spatial distortions and contrast shifts not simulated by retrospective undersampling
Tried and failed
training diffusion models from scratch on small datasets applied to image reconstruction. Outcome: overfit. Reason: training complex generative diffusion architectures from scratch with limited subject data led to severe overfitting and poor reconstruction
Multi-contrast magnetic resonance imaging with deep generative learning · UT Austin
Lost to a baseline
Direct reconstruction achieved lower MSE (0.0101) than Deep Decoder (0.1387) on synthetic sine wave data.
Lost to a baseline
When inputting late-time overdensity directly into a CNN without standard reconstruction preprocessing, the CNN failed to improve upon standard reconstruction alone.
Considered and rejected
Considered and rejected: Rejected using real MRI phase grafting onto natural images for reconstruction generalization because it limits training diversity and fails to scale to large datasets.
Generalizable deep learning based medical image segmentation · Imperial
Undersampling and non-uniform acquisition trajectories induce aliasing and streak artifacts
Sparse angular sampling, sub-Nyquist acquisition rates, and static undersampling grids fail to supply sufficient data coverage, generating severe aliasing and streak artifacts. Regular geometric sensor grids and localized scan paths consistently underperform randomized or adaptive trajectories in capturing full target fields.
Tried and failed
iterative proportional fitting applied to tomographic reconstruction from sparse 1D projections. Outcome: unstable. Reason: vulnerability to noise and inability to handle zero counts in sparse data led to reconstruction artifacts
Tried and failed
gridrec Fourier tomography reconstruction applied to sparse-view computed tomography. Outcome: worse than baseline. Reason: sparse projection sampling introduced severe noise and streak artifacts into reconstructed volumes
Mechanical Design of Selected Natural Ceramic Cellular Solids · Virginia Tech
Tried and failed
temporally static undersampling patterns applied to dynamic image reconstruction. Outcome: worse than baseline. Reason: fixed spatiotemporal sampling creates redundant aliasing compared to temporally shifted sampling trajectories
MRI techniques for quantitative and microstructure imaging · MIT
Lost to a baseline
Proposed planning method achieved slightly lower liver reconstruction precision (0.9658) than basic linear scan baseline (0.9775) due to rasterization artifacts in localized fan scans.
Integrated Path Planning System for Freehand Ultrasound Scanning · MIT
Lost to a baseline
Reconstruction and segmentation metrics on the CINE dataset were consistently lower than on SKM-TEA due to stronger aliasing artifacts induced by the VISTA undersampling pattern
Leveraging Uncertainties in Medical Prediction Systems · Publikationssystem UB Tuebingen
Considered and rejected
Considered and rejected: Rejected taking fewer than 720 photographic angles for optical tomography because reconstructions were too grainy/low-contrast for reliable segmentation
Structure and Coarsening of Foams: Beyond von Neumann's Law · Penn
Tried and failed
uniform exposure distribution across projections applied to low-dose computed tomography reconstruction. Outcome: worse than baseline. Reason: hybrid sparse normal-dose and dense low-dose sampling provides better structural priors than uniform low-dose acquisition
Data-driven X-ray Tomographic Imaging and Applications to 4D Material Characterization · Virginia Tech
Tried and failed
extreme sub-Nyquist compressed sensing reconstruction applied to ultrasound image reconstruction. Outcome: worse than baseline. Reason: severe undersampling (5% rate) caused structural distortion and drastically elevated reconstruction error
Compact modular open platform for low-cost ultrasound imaging · Imperial
Tried and failed
regular grid spatial sampling for sparse reconstruction applied to enclosed wavefield reconstruction. Outcome: worse than baseline. Reason: equidistant sampling causes spatial aliasing and coherence issues, underperforming randomized sensor placement
Tried and failed
training radiance fields from scratch on sparse data applied to dynamic scene 3D reconstruction updating. Outcome: worse than baseline. Reason: sparse updated images provided insufficient coverage in vacated regions compared to fine-tuning existing representations
Addressing Challenges in Object-Based Robot Navigation and Mapping · MIT
Tried and failed
sparse tomographic reconstruction using basis function parameterization applied to atmospheric density mapping. Outcome: data insufficient. Reason: sparse source events caused severe reconstruction errors and unrealistically high density estimates
Unifying VLF Transmitter and Lightning-driven D-Region Ionosphere Remote Sensing · Georgia Tech
Inappropriate sparsity regularization and over-regularization penalties cause artifact formation and contrast loss
Applying L1 penalties, wavelets, and excessive Tikhonov regularization disrupts natural speckle patterns, blurs sharp gradient boundaries, and amplifies background noise. Over-regularized model-based estimators fail to preserve localized peak features and yield inferior reconstructions compared to tailored baselines.
Tried and failed
default ramp denoising during 3D reconstruction applied to thick sample X-ray computed tomography. Outcome: worse than baseline. Reason: caused over-reduction and induced severe 3D noise and artefacts for thick transmission paths
The study of influence factors in x-ray computed tomography using simulation approach. · Cranfield
Tried and failed
L1 sparsity regularization in inverse problem reconstruction applied to computational 3D optical tomography. Reason: Introduced severe visual noise, background artifacts, and blurred cross-sectional edges.
Lost to a baseline
Model-based minimization acoustic reconstruction yielded lower reconstruction accuracy and smeared Bragg peak sharpness compared to iterative time-reversal.
Development of an acoustic dose-profile measurement technique for short pulse proton and ion beams · Imperial
Lost to a baseline
Simultaneous Algebraic Reconstruction Technique (SART) after 10 iterations was outperformed by GMDL-2P (achieved lower PSNR, SSIM, and higher RMSE on low-dose CT reconstruction)
Deep Learning-based CBCT Projection Interpolation, Reconstruction, and Post-processing for Radiation Therapy · DukeSpace
Lost to a baseline
Model-based ADMM reconstruction lost to deep learning-based PaDI network (yielding far less spatially informative results with high edge artifacts)
Computational Bio-Optical Imaging with Novel Sensor Arrays · DukeSpace
Tried and failed
compressed sensing error compensation without noise applied to full field modal dynamic reconstruction. Outcome: worse than baseline. Reason: compensating for non-existent errors in clean data introduced artificial artifacts and degraded accuracy
Full Field Reconstruction Enhanced With Operational Modal Analysis and Compressed Sensing for General Dynamic Loading · Virginia Tech
Tried and failed
sparse wavelet regularization using iterative shrinkage thresholding applied to ultrasound image reconstruction. Reason: Altered natural speckle patterns and improperly thresholded continuous linear intensity gradients.
Ultrasound Imaging: From Physical Modeling to Deep Learning · EPFL
Lost to a baseline
In X-ray tomography with 1% noise and active subspace dimension r = 1, TSVD (84.58% relative error) outperformed Tikhonov (21.36% relative error), DI (21.36% relative error), and DIAS (21.35% relative error) under over-regularization (alpha = 10^8) where Tikhonov reached 80.77%, DI reached 78.08%, and DIAS reached 74.29%.
Accelerating inverse solutions with machine learning and randomization · UT Austin
Considered and rejected
Considered and rejected: Rejected dense coefficient inference networks in favor of sparse coefficients because dense coefficients led to unstable operator paths and worse extrapolated reconstructions.
Manifold Learning of Neural Representations for Efficient Machine Learning Systems · Georgia Tech
Lost to a baseline
For Substantia Nigra and Dentate Nucleus ROIs, TFIR had higher susceptibility reconstruction error compared to ground truth than MEDI-SMV.
Iterative optimization schemes diverge or suffer from error accumulation across update cycles
Iterative deconvolution and phase retrieval algorithms frequently diverge or introduce spurious artifacts when exposed to laser speckle or perturbed dirty maps. Repeated inner iterations and motion correction cycles can accumulate residual errors in unsampled regions rather than improving spatial accuracy.
Tried and failed
Iterative deconvolution algorithms applied to interferometric dirty map reconstruction. Reason: alters noise covariance properties and introduces spurious residual artifacts into the map
Tried and failed
filtered backprojection and SIRT tomography reconstruction applied to continuous slab-like geometries. Reason: attenuates low-frequency components, artificially introducing non-physical through-holes in continuous slabs
Tried and failed
Iterative blind deconvolution image reconstruction applied to Coded aperture multiplexed imaging. Outcome: did not generalise. Reason: Failed to resolve periodic high-frequency patterns despite successfully reconstructing simpler quadrant features across iteration counts.
Imaging near-field compton backscattered X-rays using Pinhole and coded masks · Cranfield
Tried and failed
increasing iterative motion correction steps applied to dynamic image reconstruction. Outcome: worse than baseline. Reason: accumulating residual errors in unsampled frequency domain regions degraded reconstruction metrics
Deep learning for accelerated magnetic resonance imaging · Imperial
Tried and failed
iterative phase retrieval under unfiltered coherent illumination applied to optical image and phase reconstruction. Outcome: did not converge. Reason: laser speckle artifacts corrupted the spatial intensity measurements preventing accurate phase recovery
Laser-Based Dual-Space Microscopy · Texas Tech
Tried and failed
reducing inner iterations in inexact ADMM applied to regularized iterative image reconstruction. Outcome: worse than baseline. Reason: insufficient inner subproblem accuracy slowed overall convergence and degraded final reconstruction quality
Optimizing reconstruction and segmentation of free-breathing whole-heart CMR to enable clinical implementation · Georgia Tech
Tried and failed
alternating joint optimization and calibration applied to iterative image reconstruction. Reason: multiple alternating iterations yielded no noticeable quality improvement over a single round
On Improving the Acquisition and Reconstruction Of Spatio-Temporal Magnetic Resonance Imaging · MIT
Tried and failed
3D Fourier deconvolution with Tikhonov regularization applied to spatial scanning signal reconstruction. Reason: produced severe edge artifacts and poor agreement with measured signals
Considered and rejected
Considered and rejected: Rejected standard Plug-and-Play (PnP) iterative reconstruction because incomplete convergence analysis and iterative overhead yielded slow reconstruction.
Optimizing reconstruction and segmentation of free-breathing whole-heart CMR to enable clinical implementation · Georgia Tech
Hardware resolution constraints and lossy pre-processing representations obscure physical microstructures
Inadequate voxel or sensor resolution prevents reconstruction and segmentation pipelines from separating tightly packed micrometer-scale particle networks and fiber structures. Discarding raw radio-frequency phase information through envelope detection or position-free sensing further blurs speckle patterns and degrades volumetric accuracy.
Tried and failed
pixel-space reconstruction before downstream classification applied to heavily corrupted image classification. Outcome: worse than baseline. Reason: pixel restoration introduces artifacts and fails to recover features compared to representation-space inversion
Solving inverse problems with deep learning : from untrained to pre-trained models · UT Austin
Tried and failed
intensity thresholding and Otsu segmentation applied to phase-contrast x-ray computed tomography images. Reason: phase contrast fringe artifacts and low attenuation contrast obscured feature boundaries
Lost to a baseline
In phantom regions with smooth, low-intensity tissue relaxation, an 8-phase-cycle discrete acquisition (DPC-8) achieved a lower coefficient of variation (CoV) than fm-bSSFP min-align due to reconstruction noise artifacts dominating imperceptible bands.
Debanding in Frequency-Modulated bSSFP MRI: High-Resolution Neuroimaging · JScholarship
Tried and failed
computed tomography image segmentation applied to dense particulate composite microstructures. Outcome: data insufficient. Reason: voxel resolution was insufficient to resolve narrow matrix regions between closely packed particles
Tried and failed
training CNNs on envelope-detected data applied to ultrasound image reconstruction. Outcome: worse than baseline. Reason: envelope detection discards phase information, causing blurred speckle patterns compared to raw RF/IQ representations
Ultrasound Imaging: From Physical Modeling to Deep Learning · EPFL
Tried and failed
computed tomography and ultrasound imaging applied to micrometer-scale fibrous networks. Outcome: no signal. Reason: spatial resolution was insufficient to resolve fibers with diameters under four micrometers
Imaging, Mechanics, Construction, and Sonification of Three-Dimensional Spider Webs · MIT
Tried and failed
stainless steel dry spherical electrodes applied to low-frequency bioimpedance tomography. Outcome: worse than baseline. Reason: higher contact impedance and polarization effects at low AC frequencies compared to standard Ag/AgCl electrodes
Considered and rejected
Considered and rejected: Freehand 3D ultrasound without position sensors, rejected because lack of positional data yields inadequate volumetric reconstruction quality preventing quantitative measurements
Evaluation of tomographic 3D ultrasound for the assessment of vascular pathology · Imperial
Frequency domain gridding and coordinate interpolation inaccuracies generate geometric reconstruction errors
Digital interpolation between tilted projection trajectories and Cartesian frequency grids introduces geometric distortion and suppresses high spatial frequencies. Uncorrected non-uniform fast Fourier transforms and cone beam geometry misalignments cause phase inversions and resolution loss compared to trajectory-corrected baselines.
Tried and failed
cone beam tomosynthesis acquisition applied to volumetric spatial frequency reconstruction. Reason: cone beam geometry induced severe artifacts causing phase inversion of high-frequency components at off-center positions
Development And Evaluation Of Next Generation Tomosynthesis · Penn
Lost to a baseline
Gegenbauer reconstruction (SSIM 0.85003) and Piecewise filtered Fourier (SSIM 0.84592) lost to unfiltered Fourier baseline (SSIM 0.91128) and filtered Fourier (SSIM 0.92145) on 2D MRI datasets with estimated edges.
On the Epistemology of Gibbs Ringing Reduction Algorithm Performance · Research Repository UCD
Lost to a baseline
Piecewise filtered Fourier reconstruction (mean SSIM 0.83302) lost to unfiltered Fourier baseline (mean SSIM 0.90462) and filtered Fourier (mean SSIM 0.90831) on the 2D synthetic dataset.
On the Epistemology of Gibbs Ringing Reduction Algorithm Performance · Research Repository UCD
Considered and rejected
Considered and rejected: Rejected direct projection reconstruction/filtered back-projection due to digital interpolation inaccuracies between tilted projection vectors and Cartesian frequency grids.
Lost to a baseline
Handled interpolation of standard detector-resolution reconstructions was beaten in sharpness and resolution by native fine-grid backprojection filtering (BPF), which uniquely preserves frequencies above 3.57 lp/mm.
Modeling the Anisotropic Resolution and Noise Properties of Digital Breast Tomosynthesis · Penn
Lost to a baseline
Uncorrected full HSS reconstruction and gridding NUFFT achieved lower SSIM (0.51–0.57) on a 1H resolution phantom than BART NUFFT with GIRF correction (0.64).
Advanced spectroscopic imaging techniques for X-nuclei cardiac magnetic resonance imaging · Oxford
Left open by the authors
Problems the authors named and did not get to.
Left open
Develop iterative reconstruction methods for second-harmonic optical diffraction tomography using nonlinear beam propagation forward models to reduce missing-cone artifacts. Blocker: None
Exploring Novel Modalities for Optical Diffraction Tomography · EPFL
Left open
Apply polynomial image warping distortion correction to all projection pixels across the full field of view before CT reconstruction. Blocker: Requires raw projection image data acquired from the specific image-intensifier CT hardware setup
Improvements on low-cost high-resolution CT system Improvements on low cost high resolution CT system · Iowa State
Left open
Develop iterative or deep learning reconstruction algorithms to correct missing-cone axial elongation artifacts in tilt-angle stimulated Raman projection tomography. Blocker: None
Left open
Integrate deep learning-based projection filtering into the reconstruction pipeline and extend the method to limited-angle CBCT reconstruction. Blocker: None
Deep Learning-based CBCT Projection Interpolation, Reconstruction, and Post-processing for Radiation Therapy · DukeSpace
Left open
Investigate Deep Image Prior for mitigating multiple scattering artifacts and experimental uncertainties in large-scale optical coherence refraction tomography. Blocker: Requires multi-gigavoxel experimental OCRT dataset or specialized hardware apparatus to capture multi-angle scans
Computational 3D Optical Imaging Using Wavevector Diversity · DukeSpace
Left open
Adapt the Z-reconstruction algorithm to handle experimental artifacts, defective detector channels, and container edge motion effects. Blocker: Requires raw dual-energy cargo X-ray radiography system data with experimental artifacts and defective detector channels
Reconstructing the Atomic Number of Cargo X-ray Images using Dual Energy Radiography · MIT
Left open
Apply cone-beam refractive-index reconstruction algorithms to Optical Projection Tomography data of extended biological specimens. Blocker: Requires optical projection tomography physical experimental setup or specialized raw biological sample projection data.
Physics-based modeling and inverse problems in optical imaging · EPFL
Left open
Implement SADIR iterative reconstruction using projection domain CT data and evaluate its impact on COVID-19 classification accuracy. Blocker: Raw CT projection-domain data (sinograms) for clinical COVID-19 cases is rarely publicly accessible.
Real-Time Computed Tomography-based Medical Diagnosis Using Deep Learning · Virginia Tech
Left open
Apply the coronary biomechanics pipeline to Photon Counting Computed Tomography (PCCT) images to evaluate soft lipid detection and reduce blooming artifacts. Blocker: Access to clinical Photon Counting CT (PCCT) coronary angiography datasets.
Noninvasive assessment of coronary artery plaque vulnerability using computational solid mechanics · UT Austin
Left open
Apply synthetic data augmentation methods developed for MRI reconstruction to medical image classification tasks. Blocker: The unfinished work is a broad direction without specific architectures, target classification datasets, or validation protocols defined.
DEEP LEARNING-BASED METHODS FOR IMPROVING ACCELERATED MAGNETIC RESONANCE IMAGE RECONSTRUCTION · JScholarship
Checking a claim in this area?
We can run the same search on any method or claim. If nothing turns up, we will say so, and that proves nothing on its own.