Editable Vertex Sequences for Lumbar Spine Morphometry

Radiologists who measure the lumbar spine on MRI, for example the height of each intervertebral disc, work from 2D sagittal slices. Tracing is slow and readers disagree, and both problems get worse when the anatomy is badly degenerated. Automated tools already cut the time. Open Lumbar Spine Image Analysis reports under 10 seconds per subject against 15 minutes by hand. These tools output pixel masks, and the clinical number is computed from those masks by hidden post-processing. When the number is wrong, the radiologist cannot see which part of the boundary caused the error, and cannot cheaply fix it.

We propose predicting each vertebra and disc outline as an ordered list of boundary points, generated one point at a time, with every point conditioned on the points before it. The resulting outline is a sequence of vertices. This buys two things. First, every clinical measurement becomes an explicit function of named vertices the radiologist can see. Second, when the radiologist moves one vertex, the model regenerates the rest of that outline, and of the neighbouring structures, conditioned on the correction. A single fix therefore propagates instead of requiring many. We test whether this "edit propagation" reduces correction effort and measurement error on pathological spines. We also test whether the effort measured with simulated users on retrospective benchmarks predicts the effort of real radiologists on consecutive live scans.

Core Claim

The claim has three parts. Autoregressive polygon generation already exists for natural images; the new contribution is using its prefix conditioning as a correction mechanism for clinical morphometry. We make this three testable claims:

  • A radiologist correcting an autoregressive outline needs fewer edits to bring disc height and vertebral height ratios within the inter-reader range than a radiologist correcting a polygonized pixel mask. The gap should be concentrated in severely degenerated discs and spondylolisthesis, where pixel models are known to degrade.
  • Anchoring each displayed number to its defining vertices lowers automation bias. Automation bias here means the rate at which readers accept a deliberately wrong measurement.
  • Correction effort measured with simulated users on retrospective data can be checked against effort in a silent-mode prospective deployment, turning the retrospective-prospective gap into a measured quantity rather than an assumption.

The interpretability we care about is not a heatmap. It is the property that the output is a short, inspectable geometric program whose every token the clinician can audit and overwrite.

Relation to Prior Work

Three papers supply the method. A Unified Sequence Interface for Vision Tasks showed that masks can be emitted as token sequences of quantized coordinates. PolyFormer replaced quantized tokens with a regression decoder that outputs continuous coordinates, and separated multi-part objects with a separator token. We reuse this design, because spine morphometry needs sub-pixel geometry. Polygon-RNN++ showed that correcting a polygon's vertices cut annotator clicks by 50% and made annotation 3x faster on Cityscapes.

None of these three was evaluated on radiology morphometry or with clinicians making decisions. Polygon-RNN++'s correction loop was built for dataset annotation, not for reading a scan and reporting a number.

The medical "autoregressive" segmenter AR-Seg generates raster masks from coarse to fine scales. It motivates us, because it shows that step-by-step generation can be inspected. It differs from us in two ways. Its steps are pixel maps rather than editable geometry, and it was evaluated only offline.

Two papers show where current spine tools stop. SpineSegDiff documents that nnU-Net falls to 0.85 Dice on degenerated discs in T2-weighted SPIDER scans. Its diffusion refinement, a generator that iteratively denoises an initial nnU-Net mask, reaches 0.90. Automatic lumbar vertebra segmentation from clinical CT derives fracture diagnosis from eight boundary points per vertebra, but its hand-built pipeline over-segments L5 by up to 26%.

Together these support our premise that the clinical quantity lives on a small set of boundary points. They also raise a real alternative, direct landmark prediction, which we test below.

One paper constrains our design. Coherent concept-based explanations shows that spatially grounded concepts can raise accuracy, but it needs dense expert concept labels. We therefore do not add a concept bottleneck. The vertices themselves are the grounded intermediate representation, and they need no labels beyond the outlines already in SPIDER.

The evaluation design comes from Kottlors et al.. They ran a crossover reader study with a 3-month washout, in which AI support cut reading time by 25.7% and raised inter-reader agreement from 0.34 to 0.58, but readers only viewed predictions passively. Horng, Sontag et al. warn that retrospective accuracy does not establish prospective effect. We take both seriously by measuring interactive, not passive, use, and by comparing offline and live effort directly.

Research Program

The model. It reads a 2D sagittal T2-weighted slice with a vision transformer encoder. It emits, in a fixed anatomical order (L1 through S1, with each disc placed between its bounding vertebrae), one closed outline per structure. Each outline is a sequence of continuous coordinates, and separator tokens mark where one structure ends and the next begins.

Named vertices. A small number of vertex slots per structure are tied to anatomical meaning, such as the anterior and posterior endplate corners. These slots are trained with extra supervision from corner points derived from the SPIDER outlines. Disc height and vertebral height ratios are then closed-form functions of these slots, not of a pixel mask.

Training for edits. Training uses teacher forcing, meaning the model is fed the true previous points while predicting the next one. We add corrupted-prefix augmentation: with some probability, a ground-truth vertex in the history is displaced and then restored to its true value. This teaches the decoder to recover correctly after an edit, in the same spirit as the simulated drawing distortions used by Contour-based Interactive Segmentation.

The correction loop. At inference, a radiologist's edit to vertex $k$ fixes the history up to $k$. The decoder then regenerates everything after $k$. Latency grows with the number of vertices, but a lumbar spine needs only a few hundred points, so regeneration should stay interactive. We will measure this.

Interface. The model runs as a 3D Slicer extension, the same host environment as OLSIA, so that workflow friction is comparable.

Baselines

The naive reference point is manual tracing in the same 3D Slicer interface. OLSIA's published figure of 15 minutes per subject anchors it. Here it tells us how much of any gain is automation in general rather than our mechanism.

The strongest automatic competitor is SpineSegDiff, together with the nnU-Net it starts from. On SPIDER, SpineSegDiff reports mean Dice 0.913 against nnU-Net's 0.890, and 0.90 against 0.85 on T2-weighted degenerated discs (SpineSegDiff). We retrain both on our SPIDER split with the same patient-level folds and a matched GPU-hour budget. Morphometry is extracted from their masks by fitting polygons and detecting corners.

Beating them does not mean higher Dice; we expect roughly parity, and no reported result on Lacuna exists for polygon-sequence methods on SPIDER. It means lower disc-height error against expert measurements, with Dice no worse than 0.01 below SpineSegDiff, on the degenerated-disc and spondylolisthesis strata.

OLSIA is the deployed-tool reference. It reports mean Dice above 90% and a disc height bias of 0.02 across six external cohorts (Open Lumbar Spine Image Analysis). We run the released tool as-is on our test scans.

For interaction, the strongest competitor is a polygonized nnU-Net mask whose vertices can be dragged in the same user interface. In a second arm, the nnU-Net mask is refined with click-based interactive correction.

The key ablation is our own model with edit propagation switched off. Edits freeze the moved vertex, but the rest of the sequence is not regenerated. This isolates prefix conditioning from the benefits of a vector output and a better interface. The proposal's central claim holds only if the full model needs fewer edits than this ablation.

Why not just…?

Why not just run nnU-Net or SpineSegDiff and let radiologists drag vertices of a polygon fitted to the mask? This is the most serious threat. It delivers editable geometry without an autoregressive decoder, and it may match our model on routine spines. The difference lies in what one edit accomplishes. Moving a fitted vertex changes only that vertex, while our decoder re-derives the adjacent endplate and the neighbouring disc. The no-propagation ablation and the polygonized-nnU-Net arm settle the question. If edits-to-tolerance on degenerated discs do not fall by a clinically meaningful margin (we pre-register 30%) relative to both, the autoregressive part is not earning its cost, and we would report that.

Why not just regress the eight corner landmarks per vertebra directly, as Ghosh et al. effectively do? Disc height needs only a few points, so a keypoint model is cheaper and needs no sequence decoder. It discards the outline, however, and the outline matters in three places. Herniated or wedge-shaped discs are where corners are ambiguous. Radiologists judge corners by the shape around them. The hand-built corner pipeline in Ghosh et al. failed in exactly the low-contrast, osteophyte-laden cases. We include a heatmap landmark regressor as an extra arm in the offline comparison, stratified by pathology. If it matches us on disc height error in the hard strata, the contour is unnecessary for this task, and the proposal narrows to landmark editing.

Why not just add uncertainty or saliency maps to flag cases for review, and skip interaction? Flagging tells the reader where to look, but not what number is wrong or how to fix it. Patrício et al. document that post-hoc saliency is noisy and poorly grounded (Coherent concept-based explanations). We add a nnU-Net-plus-uncertainty-overlay arm to the reader study and measure detection of seeded errors directly.

Why not just validate retrospectively across many external cohorts, as OLSIA did? Multi-cohort validation answers whether the masks transfer across sites. It does not answer how long correction takes in practice, or whether readers wave through wrong numbers. Kottlors et al. could observe their time and agreement effects only with readers in the loop. Our silent-mode phase tests whether simulated effort predicts real effort. If it does, future work can rely more on cheap simulation, which is itself a useful negative finding for the prospective-validation argument.

Proposed Experiments

Stage 1: offline comparison. We train on SPIDER (218 subjects, T1- and T2-weighted, with pathology labels for herniation, spondylolisthesis, and other findings) using five patient-level folds. The public Mendeley spinal MRI set serves as an external test. We report:

  • Dice and Hausdorff distance per structure.
  • Absolute error in disc height index and vertebral height ratios against the SPIDER expert outlines.
  • All metrics stratified by pathology.

Simulated correction uses a scripted user who moves the worst vertex to its true position until every measurement falls within an inter-reader tolerance. The tolerance is set from a double-read subset. We compare edits-to-tolerance across all interactive arms.

Stage 2: crossover reader study. Six to eight musculoskeletal radiologists measure a fixed set of scans in each condition, separated by a washout period, following Kottlors et al. The scans are enriched for pathology. In 15% of cases, one named vertex is displaced so that a displayed measurement is outside tolerance. The outcomes are:

  • Time per subject.
  • Inter-reader agreement on disc height.
  • Measurement error against a consensus reference.
  • The acceptance rate of the seeded errors, which is our automation-bias measure.

Stage 3: silent-mode prospective deployment. Each tool runs in the background on consecutive clinical lumbar MRIs at a partner site. A study radiologist uses and corrects it without the output entering the report. We log every edit and its timing. We then compute the rank correlation between Stage 1 simulated effort and observed effort, per case and per pathology stratum.

Decisive checks. The proposal succeeds if three results hold:

  • The full model beats both the no-propagation ablation and polygonized nnU-Net on edits-to-tolerance in the degenerated strata.
  • Seeded-error acceptance is lower with vertex-anchored display than with mask overlays.
  • The simulated-to-live effort relationship is reported either way, since a weak correlation would itself show that interactive-segmentation benchmarks misstate clinical workload.