Two Clicks per Clot: Thrombus Segmentation from Routine Stroke Measurements

When a large brain artery is blocked by a clot, doctors either give a clot-dissolving drug first and then pull the clot out with a catheter (the "pharmacoinvasive" or bridging approach), or they go straight to the catheter. Choosing well depends on the clot itself: how long it is, how dense, and how much of the vessel it fills. Computer models can outline clots on CT scans, but they are trained on masks that radiologists draw voxel by voxel. That work is too slow to repeat across the many centers, trials and posterior-circulation cases where clot measurements would matter most. This note proposes training a clot segmenter without any drawn masks. The only supervision is two clicks per patient: one where contrast stops in the artery on CT angiography, and one where the clot ends. The two clicks are turned into a usable training label by exploiting a fact specific to clots. A clot sits inside a tube whose path and width can be read from the surrounding open vessel, and it shows up as a bright artery on non-contrast CT.

The core claim is narrow and testable. A model trained on 228 patients with two-click labels should come within about 0.05 volumetric Dice of the fully supervised dual-modality model of Fully Automated Thrombus Segmentation on CT on the same held-out trial. If it does, clot burden becomes cheap enough to measure across an entire randomized trial of drug-before-catheter against catheter-alone. That trial-scale measurement is the pharmacoinvasive question this note ultimately cares about.

What exists and where this differs

The fully supervised clot models motivate the proposal and also set the bar. The anterior-circulation model fuses non-contrast CT (NCCT), which shows the clot as a hyperdense artery, with CT angiography (CTA), which shows where the lumen cuts off. It reaches Surface Dice 0.78, volumetric Dice 0.62 and a 4% missed-case rate on 100 MR CLEAN-NO IV patients, and it needs dense expert masks plus a commercial detector to place its box (MIDL Short Papers). Polar-UNet handles basilar and vertebral occlusions by steering a volume of interest around the brainstem, but reaches only 0.44 Dice, again from slice-by-slice expert masks (Diagnostics). Neither paper reduces labeling cost.

Several label-efficient methods supply techniques but have never been aimed at the clot. Vessel-CAPTCHA labels vessel-containing patches with clicks and refines K-means pseudo-labels, outperforming full supervision on TOF MRA (Vessel-CAPTCHA). It was tested only on intact vessels, and its intensity clustering breaks on CT, where calcium and contrast are both bright. Active-contour weak labels for angiography segment the open vessel tree in digital subtraction angiography (DSA) at roughly a hundredth of the manual cost, but explicitly not the occluding clot (Vepa et al.).

The closest technique in spirit is RECIST co-segmentation. It turns the line measurements radiologists already record into GrabCut seeds and reaches 90.3% Dice on DeepLesion (Agarwal et al.). Its lesions are compact blobs, though. A clot is a thin tubular plug, and a line through it says almost nothing about its cross-section unless the vessel geometry is used. That geometric prior is the new ingredient here.

Two further works shape the training loop. Scribble supervision with uncertainty filtering (Can et al.) supplies the procedure of withholding uncertain pseudo-labels. Recursive self-training for lung cysts (Zhang et al.) peaked at its second round, 75.87% Dice, and fell at the third, 74.94%. That constrains the design: self-training rounds must be stopped on held-out evidence and not run open-ended.

Method

The annotator opens co-registered NCCT and CTA. They place a proximal click at the contrast cut-off and a distal click where contrast reappears through collateral filling, or where the hyperdense sign ends. The time per case will be measured, not assumed.

A pseudo-label is then built in three steps. First, a minimal-cost path along a CTA vesselness map connects the two clicks; this gives the centerline. Second, the lumen radius is estimated from the patent segment just upstream and from the mirrored contralateral artery, which together define a tube $T$ around the centerline. Third, inside $T$, voxels whose NCCT density exceeds a patient-specific threshold, set from the contralateral vessel, are labeled clot. Tube voxels below that threshold are left unlabeled. Everything outside a dilated tube is labeled background.

A dual-modality U-Net with the weighted-sum fusion of the MIDL model is trained on these labels with $$\mathcal{L} = \sum_{v \in C} \mathrm{CE}(p_v, y_v) + \lambda \sum_{v \notin T} p_v,$$ where $C$ is the set of confidently labeled voxels, $p_v$ is the predicted clot probability at voxel $v$, $y_v$ is its pseudo-label, and the second term penalizes predicted clot outside the tube, weighted by $\lambda$. The model then relabels the training set. One or two self-training rounds follow with Monte Carlo dropout filtering in the manner of Can et al., and the stopping round is chosen on a 10-case dev set that has full masks.

Because the clicks define the clot's location, the method does not need the commercial detector. At inference it can use either the clicks or a learned localizer.

Research Program

The work has three stages. The first is pseudo-label quality. On the 228 MR CLEAN and global-cohort training patients, whose expert masks already exist and are used here only for measurement, we compare tube-prior pseudo-labels against GrabCut-from-line pseudo-labels directly. The second is the label-cost curve. We train with two-click labels on all 228 patients plus 0, 10, 25 or 50 full masks, and plot test Dice against total annotator minutes. This mirrors how Vepa et al. mixed weak labels with a small share of manual ones. The third is transfer to the posterior circulation, where data are scarce and the tube prior should matter most, because tortuous anatomy defeats unconstrained models.

The pharmacoinvasive stage depends on access. MR CLEAN-NO IV randomized intravenous thrombolysis before thrombectomy against thrombectomy alone, so clot length, volume and density measured automatically at baseline can be tested as modifiers of the treatment effect on reperfusion and 90-day outcome. The 100-patient test set is far too small for this; the point of two-click labeling is that the full trial cohort can be annotated in hours. We treat this analysis as exploratory and pre-register the biomarkers.

It is related to work showing that tabular post-thrombectomy models gain discrimination over time, from AUC 0.824 at admission to 0.945 at discharge, but use no imaging (Brain, 2022). Here we add image-derived clot features to the admission model rather than relying on later clinical data.

Baselines

The naive reference point is the tube itself, or a fixed-radius cylinder around the click path, used directly as the prediction with no learning. If the trained model does not clearly beat this, the network adds nothing beyond geometry.

The strongest existing method is the fully supervised weighted-sum dual-modality U-Net, which reports Surface Dice 0.78 (95% CI 0.73–0.83), volumetric Dice 0.62 and a 4% missed-case rate on MR CLEAN-NO IV (MIDL Short Papers). We retrain it on the same 228 masked patients with the same architecture, augmentation and tuning budget. That gives a same-pipeline upper bound and avoids comparing against a number produced under unknown conditions. Beating it is not the goal. The weak model counts as successful if it reaches volumetric Dice of at least 0.57 and Surface Dice within 0.05 of the retrained model, without raising the missed-case rate above roughly twice 4%.

For the posterior circulation, the reference is Polar-UNet at 0.44 Dice and 0.65 precision (Diagnostics). Matching it with two-click labels would already be a substantial cost win, and exceeding it would suggest the tube prior helps where anatomy is hard.

The strongest weak competitor is RECIST-style line supervision: GrabCut seeds from the click segment, the same U-Net, and the same self-training (Agarwal et al.). A second weak competitor is scribble supervision with random-walk expansion, a conditional random field and dropout filtering (Can et al.), given an equal annotation time budget.

Ablations remove one ingredient at a time: the lumen tube, the NCCT density gate, the outside-tube penalty, and the uncertainty filter. The tube ablation is the one that isolates the proposal's key ingredient.

Why not just…?

Why not just prompt a segmentation foundation model such as SAM with the same two clicks? This is a fair zero-training competitor, and we run it, slice by slice and in 3D variants, on the same test set. No reported result for SAM-family models on thrombus segmentation was found on Lacuna. The indirect evidence points against it. Dropping NCCT from the MIDL model cut Surface Dice from 0.75 to 0.51, and the clot is a small, low-contrast target whose boundary is defined jointly by the two scans. Single-image prompting cannot use the second scan. If prompted SAM matches the trained model, the trained model is unnecessary and we will report that.

Why not just train fully supervised on the masks that MR CLEAN already has? For the anterior circulation in these trials, that is what one should do, and it serves as our upper bound. The problem is everywhere else. Posterior-circulation cohorts, new centers and full trial cohorts lack masks, and producing them took slice-by-slice expert work in every paper here. The label-cost curve answers this objection directly. If 25 full masks plus clicks on everything beat two-click labels alone at equal annotator time, the recommendation becomes "a few masks plus clicks," and we will report that recommendation.

Why not just seed GrabCut from the line, as RECIST methods do? This is the most serious threat, because it is the same idea without the vessel prior. It is run as a baseline under identical training. If it matches the tube-prior model within noise on both circulations, the geometric contribution is not real and the claim shrinks to "line supervision transfers to clots."

Why not just threshold hyperdense voxels? Vepa et al. found that fully automatic active contours were not usable without human candidate selection. Zhang et al.'s unsupervised K-means graph-cut teacher reached 74.67%, and the semi-automated baseline in that work reached 62.64%. On CT, calcified plaque and contrast share intensities (Vessel-CAPTCHA). The naive tube-plus-threshold baseline tests this directly.

Proposed Experiments

Training uses 129 MR CLEAN and 99 global-cohort patients with two-click labels. Testing uses the 100 MR CLEAN-NO IV patients, and posterior-circulation transfer uses the 187-patient MR CLEAN Registry cohort. All of these are trial and registry data that require collaborator access. Metrics are volumetric Dice, Surface Dice, missed-case rate, volume intraclass correlation (Polar-UNet's reported ICC was about 0.41), and annotator minutes per case.

Three outcomes decide whether the idea holds. The first is whether the two-click model comes within 0.05 Dice of the retrained supervised model. The second is whether removing the tube prior costs more than that margin, especially in the posterior circulation. The third is whether pseudo-label Dice against the existing expert masks predicts the gap to full supervision. If the first holds and the second fails, the contribution is cheap labeling but not the geometric prior. If both hold, the trial-scale analysis of clot burden as a modifier of the benefit of thrombolysis before thrombectomy becomes affordable, and that analysis is the next experiment.