Fine-Grained TAVR Pseudo-Labels

• The paper presents a method to automatically derive TAVR-specific substructures from coarse CT segmentations using distance-based rules and plane fitting.
• It details a multimodal framework that induces latent risk-conditioned masks to guide report generation and reduce hallucination in clinical outputs.
• Methodologies achieve notable improvements in segmentation metrics (e.g., Dice and IoU) by incorporating focal and skeleton recall losses over fine-grained structures.

Fine-grained TAVR-relevant pseudo-labels are automatically generated supervision signals that encode sub-anatomical, region-level, or token-level structure needed for transcatheter aortic valve replacement planning without requiring dense manual annotation. In current TAVR literature, the term refers chiefly to two complementary constructions: rule-based anatomical pseudo-labels derived from coarse CT segmentations for segmentation and measurement, and risk-conditioned grounding masks induced from global risk labels plus report supervision for multimodal reasoning and hallucination control. Together, these approaches turn weak or coarse supervision into clinically targeted fine-grained signals over annulus, valve, aortic root, LVOT, access vessels, and report-relevant evidence regions (Zöllner et al., 22 Jul 2025, Lu et al., 25 Jun 2026).

1. Definition and conceptual scope

TAVR planning is driven by structures and decisions that are finer than the label spaces of many public datasets and coarser than full dense expert annotation can economically support. One line of work therefore starts from coarse anatomical segmentations such as aorta, left ventricle, and iliac arteries, then derives TAVR-specific substructures such as aortic root, aortic valve, and annulus. Another line starts from multimodal patient-level risk labels and clinician-authored reports, then induces latent spatial masks and token-level grounding supports inside a report generator. In both cases, the pseudo-label is “pseudo” because it is generated automatically rather than manually delineated at the target granularity (Zöllner et al., 22 Jul 2025, Lu et al., 25 Jun 2026).

Source	Pseudo-label type	Primary use
"Semantic Segmentation for Preoperative Planning in Transcatheter Aortic Valve Replacement" (Zöllner et al., 22 Jul 2025)	Aortic root, aortic valve, annulus, and retained iliac classes derived from coarse CT labels	Supervised semantic segmentation for TAVR planning
"TAVR-VLM: Risk-Conditioned Causal Grounding for Hallucination-Resistant Report Generation" (Lu et al., 25 Jun 2026)	Global risk mask $M_{\text{global}}$, support mask $B_{\text{global}}$, and token-level masks $\tilde{M}_t$ over CT tokens	Multimodal grounding, interpretability, and hallucination reduction

The clinical motivation is consistent across both formulations. TAVR planning depends on annulus shape and size, aortic root geometry, LVOT, vascular access, calcifications, and coronary ostia relations. Yet available supervision is often global, coarse, or text-level rather than region-level. Fine-grained pseudo-labels are therefore used to bridge the gap between clinically relevant structure and available supervision.

2. Geometric derivation from coarse anatomical segmentations

The anatomical formulation is built on the observation that public CT datasets such as TotalSegmentator provide coarse labels, whereas TAVR planning requires finer substructures. The enriched label set comprises aorta, left ventricle, aortic root, aortic valve, annulus, iliac artery left, and iliac artery right, and spans 578 CTs. Here, aortic root, aortic valve, and annulus are pseudo-labels, while aorta, left ventricle, and iliac arteries are retained from the original dataset (Zöllner et al., 22 Jul 2025).

The aortic valve pseudo-label is extracted from the aorta by a distance rule at the LV–aorta interface. If $A \subset \mathbb{Z}^3$ denotes aorta voxels and $L \subset \mathbb{Z}^3$ left-ventricle voxels, then the valve label is

$$V = \{ v \in A \mid d(v, L) \le 3\ \text{voxels} \}.$$

The annulus pseudo-label is defined on the ventricular side of the interface:

$R = \{ l \in L \mid d(l, A) = 1\ \text{voxel} \}.$

These rules are explicitly designed so that the valve becomes a curved aortic-side zone near the ventricle, whereas the annulus becomes a thin ring-like slice approximating the annular plane used for prosthetic sizing.

The aortic root pseudo-label is generated by a more elaborate geometric procedure. First, annulus voxels are used to fit a plane by least squares. Parallel planes are then moved along the normal direction, and for each plane the intersection area with the aorta is measured. The paper reports an empirical pattern with a first local maximum corresponding to the sinuses of Valsalva, followed by a local minimum around distance 25 voxels, interpreted as the end of the root before the ascending aorta. After moving-average smoothing, the root is defined as the aortic segment between the annulus plane and this local minimum plane.

This construction makes the pseudo-labels explicitly TAVR-oriented rather than merely anatomical. The refinement is motivated by tasks such as annulus measurement, prosthesis sizing, root geometry assessment, and transfemoral access evaluation. The key feature is that all of this structure is induced from existing coarse segmentations without additional manual voxelwise delineation of the target substructures.

3. Segmentation models trained on TAVR-specific pseudo-labels

Once generated, the pseudo-labels are treated as training targets in standard supervised semantic segmentation. The evaluated architectures are 3D U-Net, 3D V-Net, and Swin UNETR, all implemented with MONAI. The dataset split is 378 / 100 / 100 CTs for train/validation/test, and training runs for 350 epochs on an NVIDIA RTX 2080 Ti with 12 GB memory (Zöllner et al., 22 Jul 2025).

The paper emphasizes that loss design is especially important because valve, annulus, and iliac arteries are small or thin structures. The baseline is Dice plus cross-entropy with weighting $0.25$ Dice and $0.75$ cross-entropy:

$$\mathcal{L}_{\text{DiceCE}} = 0.25 \cdot \mathcal{L}_{\text{Dice}} + 0.75 \cdot \mathcal{L}_{\text{CE}}.$$

Focal loss is added for class imbalance with focusing parameter $B_{\text{global}}$0, and skeleton recall loss is used to preserve connectivity of thin structures. The proposed focal skeleton recall loss combines focal weighting with recall on class skeletons, and the final segmentation objective is

$B_{\text{global}}$1

Quantitatively, the best validation result with Swin UNETR rises from 81.78% mean Dice with DiceCE only to 83.50% with focal plus focal skeleton recall. On the test set, Swin UNETR improves from 81.86% mean Dice and 72.86% mean IoU with DiceCE to 83.20% mean Dice and 74.79% mean IoU with $B_{\text{global}}$2. The abstract summarizes this as a +1.27% Dice increase. Per-class Dice gains are concentrated in the clinically small or thin structures: aortic root 84.63 to 86.44, valve 81.78 to 83.28, annulus 53.80 to 55.23, iliac artery left 79.81 to 81.14, and iliac artery right 81.71 to 83.43 (Zöllner et al., 22 Jul 2025).

These results frame fine-grained pseudo-labels as more than a data-generation convenience. They define a benchmark and supervision layer that make relevant structures measurable in CT scans, even though the paper does not yet report downstream annulus diameters, coronary ostia distances, or access-vessel measurements against manual clinical measurements. The released resource is intended as a benchmark and as a starting point for TAVR-oriented segmentation and planning pipelines.

4. Risk-conditioned latent pseudo-labels for multimodal grounding

A second formulation appears in TAVR-VLM, where pseudo-labels are not explicit anatomical classes but latent risk-conditioned masks over CT tokens. TAVR-VLM formulates TAVR planning as a multimodal reasoning problem with inputs

$B_{\text{global}}$3

where $B_{\text{global}}$4 is 3D cardiac CT, $B_{\text{global}}$5 is 2D echocardiography, and $B_{\text{global}}$6 is a clinical/pathological biomarker vector including variables such as STS score, creatinine, EF, valve gradient, and comorbidities. The outputs are a procedural risk level $B_{\text{global}}$7 and a structured report $B_{\text{global}}$8 containing findings and recommendations. The benchmark is $B_{\text{global}}$9, a 1,482-patient cohort with 200 test cases carrying expert ROIs for spatial-grounding evaluation only (Lu et al., 25 Jun 2026).

The pseudo-label mechanism is organized as a “Risk $\tilde{M}_t$0 Region $\tilde{M}_t$1 Word” pathway. A fusion network predicts a risk distribution $\tilde{M}_t$2, which is projected through a learnable risk prototype matrix $\tilde{M}_t$3 to form the causal bottleneck

$\tilde{M}_t$4

Cross-attention from this bottleneck to dense CT tokens $\tilde{M}_t$5 produces the global risk mask

$\tilde{M}_t$6

This mask is a continuous distribution over CT patches and functions as a risk-conditioned region-importance pseudo-label.

A discrete support mask is then obtained by top-$\tilde{M}_t$7 selection:

$\tilde{M}_t$8

yielding a binary support region over CT tokens. During autoregressive report generation, clinically salient tokens $\tilde{M}_t$9 receive raw token-level attentions $A \subset \mathbb{Z}^3$0, which are support-projected into

$A \subset \mathbb{Z}^3$1

with containment

$A \subset \mathbb{Z}^3$2

In this formulation, $A \subset \mathbb{Z}^3$3, $A \subset \mathbb{Z}^3$4, and $A \subset \mathbb{Z}^3$5 are the fine-grained pseudo-labels. They are not manually annotated; they are induced from global risk supervision, language modeling, and causal consistency constraints.

This design links high-level procedural risk to local anatomical evidence. The paper gives examples of report content such as “Severe aortic valve calcification.”, “Annulus diameter: 24.6 mm; recommend 26 mm valve.”, “High risk of paravalvular leak; consider post-dilation.”, and “Femoral access feasible.” It also states that risk labels reflect VARC-3-like procedural complication risks including annular rupture, coronary obstruction, PVL, and conduction disturbances. The induced masks therefore operate as task-conditioned evidence maps rather than generic anatomical segmentations.

5. Hallucination control, spatial grounding, and interpretability

In TAVR-VLM, hallucinations are defined as mentions of clinical entities without supporting multimodal evidence. The support-projected causal consistency loss is the mechanism that turns latent masks into active constraints:

$A \subset \mathbb{Z}^3$6

The first term aligns token-level grounding with the global risk distribution inside the allowed support; the second penalizes leakage outside that support. StopGrad makes the path strictly top-down: risk drives masks, and tokens adapt to masks rather than expanding them (Lu et al., 25 Jun 2026).

The full objective is

$A \subset \mathbb{Z}^3$7

with $A \subset \mathbb{Z}^3$8 empirically set in Fig. 2a. The paper states that higher values improve mIoU and reduce hallucination, but values that are too large make the generator rigid.

On $A \subset \mathbb{Z}^3$9, the model achieves an AUROC of 0.896, a CIDEr of 0.936, and a hallucination rate of 8.1%. Spatial grounding is evaluated by mIoU against expert ROI masks on the 200-case subset, and the full model reaches mIoU 0.624. The ablation table clarifies the role of pseudo-labels. Without purification ($L \subset \mathbb{Z}^3$0), hallucination rate rises from 8.1% to 18.4% and mIoU falls from 0.624 to 0.485. Without the causal loss ($L \subset \mathbb{Z}^3$1), hallucinations rise to 16.1% and mIoU drops to 0.350. Without stop-gradient, mIoU falls to 0.412 and hallucinations reach 14.3%. Qualitative overlays show that for “significant leaflet calcification with LVOT interaction,” the learned masks focus on calcified cusps and LVOT, whereas the baseline without R-CGA attends to diffuse or irrelevant regions.

The paper also makes an explicit downstream claim about reuse. Global risk masks and constrained token masks can be directly used as fine-grained pseudo-labels for future segmentation or detection models, and can also serve as teacher signals for distillation. This suggests a second life for grounding masks beyond report generation itself.

6. Limitations, misconceptions, and future directions

A recurring misconception is to treat pseudo-labels as equivalent to dense expert ground truth. The TAVR papers do not make that claim. In the segmentation setting, valve and annulus labels are heuristic constructions from voxel-distance rules and root boundaries are defined by a cross-sectional extremum heuristic; there is no manual correction of these target labels, and annulus Dice remains in the low- to mid-50% range. In the grounding setting, expert ROIs exist only for a 200-case test subset and are used only for evaluation; the learned masks are otherwise latent and task-conditioned rather than directly supervised (Zöllner et al., 22 Jul 2025, Lu et al., 25 Jun 2026).

The two strands also have different failure modes. Geometry-derived pseudo-labels are limited by the quality and style of the underlying coarse segmentations and do not yet include coronary ostia, calcifications, femoral arteries beyond the iliacs, or detailed LVOT subregions. The paper explicitly states that these structures likely require manual expert annotation. Risk-conditioned latent pseudo-labels are limited to CT token or patch granularity, are biased toward entities explicitly mentioned in reports, and are optimized for risk prediction plus report generation rather than for general anatomical segmentation. TAVR-VLM also reports no clinician user study beyond mIoU and hallucination metrics (Zöllner et al., 22 Jul 2025, Lu et al., 25 Jun 2026).

Even with these limitations, the two formulations are complementary rather than redundant. Geometry-derived labels provide explicit anatomical classes for segmentation and measurement. Risk-conditioned masks provide task-adaptive evidence localization tied to procedural risk and language generation. A plausible implication is that future TAVR systems may combine both: explicit pseudo-labels for measurable anatomy and latent pseudo-labels for decision-specific evidence selection. The immediate research directions already identified in the source papers include more anatomical labels, integration with measurement pipelines and clinical outcomes, refinement of valve, annulus, and root definitions, incorporation of calcification and coronary ostia information, and reuse of grounding masks for segmentation, complication localization, and automatic measurement annotation.