Searching Segmenters for Fast Spine MRI by Following the Aliasing

Spine MRI exams are long. A clinical protocol can take 30 to 60 minutes (Puy et al.), and the main way to shorten them is to skip measurements in the scanner's frequency domain ($k$-space, the space of spatial frequencies the scanner fills line by line). Skipping lines does not blur the image the way a low-resolution scan would. It folds displaced "ghost" copies of the anatomy across the image. For the spine this matters in a specific way: the spine is a chain of near-identical vertebrae, so a ghost of one vertebra lying on top of its neighbor is close to the worst possible artifact for a model that has to count and name vertebrae. This proposal asks whether a segmentation network built specifically for these folded images, and found by architecture search, can produce correct vertebral labels and boundaries from much shorter scans than current spine models tolerate. The image is never reconstructed for a human reader along the way.

The core idea fits in two sentences. Regular undersampling folds the image along one known axis at known offsets, so the operations a network needs in order to "unfold" it, such as long-range mixing along that axis or a learned frequency-domain filter, can be offered to an architecture search as candidate building blocks. A memory-efficient search, one that runs only a sampled subset of operations and channels at each training step, lets that search see whole 3D lumbar volumes rather than small patches, which is necessary because vertebra identity is decided by global context. What would be new is the claim that the best segmenter architecture changes with acceleration in a way the physics predicts. A hand-designed network trained on the same folded data should not recover that change.

Core Claim

If a fully sampled image $x$ is measured on a Cartesian grid with mask $M$ that keeps every $R$-th phase-encoding line plus a fully sampled low-frequency center, the network input is the zero-filled image

$$x_0 = F^{-1} M F x,$$

where $F$ is the 2D Fourier transform in the sagittal plane. Outside the center band, $x_0$ is approximately the true image plus $R-1$ copies shifted by multiples of $\mathrm{FOV}/R$ along the phase-encoding axis, where $\mathrm{FOV}$ is the field of view. The resulting confusion is non-local but structured. A pixel's ghost sits a fixed, known distance away along a single axis. Two predictions follow. First, at high $R$, searched architectures will put axis-specific long-range operations, or frequency-domain operations, at specific depths, and they will do this more often as $R$ grows. Second, the errors that matter clinically, namely the "one-label shift" in which every vertebra is named one level off (VerSe) and the loss of thin posterior elements, will grow with acceleration faster than overall Dice falls. Ranking methods by Dice alone would therefore hide the failures that matter. The proposal is successful if the searched networks hold identification rate and posterior-element Dice at accelerations where strong fixed baselines break.

Relation to Prior Work

Two lines of work supply the search machinery. ProxylessNAS runs one sampled operation per edge so that memory does not grow with the number of candidates. SCNAS applies a related two-operation Gumbel-Softmax sampling to 3D medical segmentation, but only on fully sampled images, with a fixed macro-architecture and about a day on four V100s. PC-DARTS instead routes only a fraction of channels through candidate operations. It also constrains our design: accuracy dropped sharply going from a 1/4 to a 1/8 channel fraction, so we stay at 1/4 and take further savings from operation sampling. MS-NAS and Thrifty NAS show that partial-channel search finds compact segmenters, but only in 2D and on clean soft-tissue images. We reuse operation sampling and partial channels together. The new elements are the input (folded, undersampled volumes), the candidate operations (derived from the folding geometry), and the use of the saved memory (enlarging the training crop to the whole lumbar volume).

Two papers connect search or segmentation to acceleration, and each leaves a gap. EMR-NAS searches a 2D reconstruction network for image fidelity, measured by PSNR and SSIM, on cardiac and brain data. We do not reconstruct and do not optimize fidelity. Joint calibrationless reconstruction and segmentation shares one hand-designed encoder between an unrolled reconstruction and a segmentation decoder on brain MRI. It motivates this proposal by arguing that sequential pipelines pass reconstruction artifacts into segmentation, and it is a baseline here. Optimal undersampling patterns for pathology localization (IGS) optimizes the mask for a fixed downstream segmenter and reports up to 12% Dice gain at 16x over standard masks. That is the mirror image of our question, since IGS varies the mask with the network fixed and we vary the network with the mask fixed. Huijben et al. supply a differentiable mask layer, which we use only in a secondary experiment.

On the spine side, SPINEPS (nnU-Net semantic segmentation followed by sliding-window instance labeling) and SpineSegDiff (diffusion refinement of an nnU-Net mask) set the accuracy bar on fully sampled scans, and neither has been characterized under undersampling. SAU-Net adds hand-designed attention between adjacent slices to a 2D network. That is exactly the kind of fixed, local context design whose fit to non-local folding we test.

Research Program

We will build a retrospective undersampling pipeline for sagittal lumbar MRI. We will then define a search space with physics-motivated operations, run a memory-efficient one-shot search at each acceleration factor, and analyze both the accuracy and the structure of the resulting architectures.

The pipeline simulates single-coil Cartesian undersampling from SPIDER's magnitude T1- and T2-weighted volumes at $R \in \{4, 8, 12, 16\}$, using equispaced and random masks with a fixed low-frequency center. This simulation lacks real phase and coil information. We state that limitation up front and check its consequences in a separate experiment, described below.

The search space is a 3D encoder-decoder whose cells choose among ordinary operations (3D convolutions of several sizes, dilated convolutions, anisotropic convolutions with kernels like 1×3×3) and three folding-aware operations:

  • a 1D convolution along the phase-encoding axis whose dilation equals $\mathrm{FOV}/R$, so it reaches exactly the ghost locations;
  • axial self-attention along the phase-encoding axis;
  • a learned pointwise filter applied in the Fourier domain of the feature map, followed by a projection that restores the measured $k$-space lines (a feature-level data-consistency step).

Unlike SCNAS, the search also chooses how many downsampling stages to use.

Supernet training samples two operations per edge and routes a quarter of the channels. We will measure peak memory against full DARTS-style search and spend the savings on crop size, aiming for the entire lumbar volume on a single 48 GB GPU. The search loss is Dice plus cross-entropy. An instance-labeling head predicts vertebra identity, so that the search is penalized for one-label shifts and not only for poor voxel overlap. After each search, the derived network is retrained from scratch. We then analyze where folding-aware operations end up as a function of $R$ and of mask type. Random masks produce incoherent, noise-like artifacts rather than structured ghosts, so the physics account predicts that the preference for ghost-reaching operations should weaken or disappear under random masks. That is a falsifiable control.

Baselines

The naive reference point is the published SPINEPS model run unchanged on zero-filled undersampled SPIDER volumes. It shows how badly an off-the-shelf spine segmenter fails as scans get shorter. On fully sampled SPIDER, SPINEPS reports an instance-wise vertebra Dice of 0.929 (SPINEPS), and SpineSegDiff reports mean Dice of 0.913 against nnU-Net's 0.890 (SpineSegDiff). These are ceilings, not undersampled baselines. No reported result was found on Lacuna for any spine segmenter on undersampled spine MRI, so every undersampled number in this study will be produced by us under one protocol.

The strongest competitor is nnU-Net retrained at each acceleration on exactly the same simulated zero-filled volumes, with the same instance-labeling stage as SPINEPS. It is the strongest because it adapts its own preprocessing and network configuration, and because training on folded data already lets a fixed network learn to ignore ghosts. If architecture does not matter, this baseline will show it. To keep the comparison fair, nnU-Net receives its standard five-fold training plus extra GPU-hours equal to our total search cost, spent on extra epochs or larger patches. All methods use the same splits and the same masks. Beating it means a statistically significant improvement, by paired bootstrap across subjects, in vertebra identification rate and posterior-element Dice at 8x and above, without losing more than one Dice point on vertebral bodies.

The sequential-pipeline baseline is a fastMRI-style reconstruction U-Net (fastMRI) trained on the same simulated data, followed by nnU-Net trained on its outputs. The joint baseline reimplements the fixed shared-encoder design of Pramanik and Jacob in the single-coil setting. The efficiency baseline is SAU-Net, retrained on folded inputs.

Three ablations isolate the proposal's ingredients. The first removes the folding-aware operations from the search space while keeping everything else, which tests whether the physics-motivated operations matter. The second runs the identical search on 96³ patches instead of whole volumes, which tests whether spending memory on context is what prevents label shifts. The third is a hand-built network that inserts the most-selected folding-aware operation at every stage. It tests whether search adds anything beyond identifying a useful operation type.

Why not just…?

Why not just retrain nnU-Net on undersampled images? This is the real threat. It may come close, because a fixed network with enough data can learn to suppress structured ghosts. Nothing in the file shows that it fails. IGS's 12% gain at 16x comes from changing the mask, not the network. The nnU-Net baseline above is designed to settle the question. If it matches the searched networks on identification rate and posterior-element Dice at every $R$, we will report that architecture is not the bottleneck under acceleration. That would still be a useful negative result, because it would redirect effort toward mask design.

Why not reconstruct first and then segment? Reconstruction losses reward smooth, low-error images, and the file notes that these can wash out fine structure. Thin posterior elements are already the weakest structures even on clean scans: SPINEPS reports 0.838 Dice for spinous processes against 0.920 for vertebrae. Pramanik and Jacob report that joint training beat sequential reconstruction networks at 6x on brain data. Our sequential baseline tests whether this holds for bone, and our per-structure breakdown shows where any loss occurs.

Why not use a hand-designed context module like SAU-Net's? SAU-Net's attention looks at neighboring slices through the plane. Cartesian folding displaces ghosts within the plane, along the phase-encoding axis, by a distance that depends on $R$. A fixed local design therefore points in the wrong direction and covers the wrong distance. The hand-built "insert the winning operation everywhere" ablation is the fair version of this alternative. If it matches the searched network, the honest conclusion is that the contribution lies in the operation, not the search.

Why not just optimize the mask, as IGS does? Mask optimization is complementary, and it may deliver larger gains per unit of effort. In a secondary experiment we learn a probabilistic line mask with the layer of Huijben et al., both with nnU-Net and with our searched network. If the learned mask with nnU-Net beats our fixed-mask searched network, that ordering will be reported as the main practical finding.

Proposed Experiments and Decisive Checks

The primary benchmark is SPIDER (218 subjects, sagittal lumbar T1- and T2-weighted), with subject-level train, validation, and test splits. If access allows, SpineSegT2W (195 T2w scans) serves as a second test set. Metrics are Dice per structure (vertebral body, posterior elements, intervertebral discs), Hausdorff distance, VerSe-style vertebra identification rate, and one-label-shift rate per subject. Efficiency is reported as peak training memory, search GPU-hours, and inference time.

Three checks decide the outcome. The first is the accuracy-versus-acceleration curve. The claim holds only if the searched networks beat retrained nnU-Net on identification and posterior-element Dice at 8x or above. The second is the structural prediction. Folding-aware operations should be selected more often as $R$ increases under equispaced masks, and noticeably less often under random masks. If the selections show no such dependence, the physics account is wrong, even if accuracy improves. The third addresses simulation realism. Our data are simulated from magnitude images, so we repeat the method ranking on the fastMRI knee raw multi-coil $k$-space task used by IGS, comparing simulated-from-magnitude undersampling with true raw-data undersampling. If the ranking of methods flips between the two, conclusions drawn from the spine simulations are not trustworthy, and we will say so before any claim about real shortened spine scans.