Uncertainty-Aware Compositional Localization and Placement Assessment of Catheters and Tubes in Chest X-Rays

Paper: arXiv:2608.11288v1 [eess.IV], 11 Aug 2026

The paper addresses the safety-critical yet tedious and error-prone task of assessing catheter and tube placement on chest X-rays. Current deep learning methods either classify placement globally—losing track of which device is where—or segment all devices into a single mask, making per-device assessment impossible when catheters overlap. The authors introduce UCompCXR, a compositional framework that detects local catheter fragments, associates them into device instances via graph-based clustering, fuses per-fragment tip predictions through precision-weighted Gaussian estimation, and classifies placement per device.

On the RANZCR CLiP dataset (30,083 images, 5-fold patient-level cross-validation with bootstrap confidence intervals), UCompCXR detects 26% more devices than a strong multi-task baseline sharing the same MobileNetV3 backbone, with 75% fewer false positives and well-calibrated tip uncertainty (95% coverage = 0.948). The aggregate tip error rises—but only because the model finds devices the baseline misses entirely, especially nasogastric tubes. On matched devices, catastrophic localization failures drop substantially. At 2.27M parameters in a single forward pass, the model is deployable on resource-constrained clinical hardware.

Catheters and tubes are ubiquitous in hospitalized patients: endotracheal tubes (ETTs) secure airways; nasogastric tubes (NETs) deliver enteral nutrition; central venous catheters (CVCs) and Swan-Ganz catheters enable medication delivery and hemodynamic monitoring. After placement, a chest X-ray is almost always ordered to confirm position—by some estimates, 50–70% of ICU chest films exist solely for this purpose. The stakes are real: an ETT pushed too far can collapse a lung, a CVC tip drifting into the right atrium invites arrhythmias, and a coiled NET that never reaches the stomach does nothing but increase aspiration risk.

Reading these films is slow and surprisingly error-prone. Roughly 1–3% of central lines need repositioning after initial imaging, and delayed detection of malposition tracks directly with complication rates. An ICU chest X-ray often contains three or more overlapping devices, each requiring independent assessment.

Limitations of existing approaches:

• Global classification methods (RANZCR CLiP challenge, Pham et al.) predict image-level labels like ETT Abnormal but cannot point to the offending device.

• Segmentation approaches (Yi et al., Lian et al.) trace devices but lump all instances of a given type into one mask; when two CVCs overlap, there is no way to assess them individually.

• Neither approach recovers the full compositional structure: a variable number of device instances, each with a traced path, a tip location carrying spatial uncertainty, and a placement label.

1. Compositional catheter representation: A fragment-based decomposition in which local catheter segments are detected, associated into device instances via graph-based clustering, and individually assessed for placement status. This is the first method that jointly performs instance-level device detection, tip localization with calibrated uncertainty, and per-device placement classification in a single forward pass.

2. Precision-weighted Gaussian tip fusion: A heteroscedastic tip prediction formulation in which each fragment predicts a tip location with an associated covariance, and per-device tips are obtained via iteratively-reweighted least squares (IRLS) with a Huber kernel. This yields well-calibrated tip uncertainty (95% coverage = 0.948).

3. Noisy-OR weak supervision: A noisy-OR formulation that bridges per-device placement predictions to image-level supervision, enabling end-to-end training on the large pool of weakly-labeled images that lack device trace annotations.

4. Comprehensive evaluation: Rigorous evaluation on the RANZCR CLiP dataset using patient-level 5-fold cross-validation with 2,000-iteration bootstrap confidence intervals. The analysis reveals a nuanced detection–localization trade-off.

The RANZCR CLiP dataset provides two tiers of supervision:

• Strong annotations (30.2% of images): traced device paths as ordered polylines with per-device family labels.

• Weak annotations (all images): 11 binary image-level placement labels spanning three device families—ETT (Normal, Borderline, Abnormal), NET (Normal, Abnormal, Incompletely Imaged), and CVC (Normal, Abnormal, Borderline, with separate labels for four CVC subtypes).

Both the baseline (B2) and UCompCXR (B7) share a lightweight feature extractor. The encoder is MobileNetV3-Small, producing multi-scale feature maps at strides 4, 8, 16 relative to the input resolution of 768 × 768 pixels. Features are fused through a CompactFPN (Compact Feature Pyramid Network) using depthwise separable convolutions and channel reduction. The output is a unified feature tensor at stride 4.

The baseline (1.99M parameters) attaches three parallel heads:

1. Line segmentation head: Predicts a per-pixel binary mask for catheter presence, supervised with binary cross-entropy, soft Dice, and clDice (topology-preserving loss).

2. Tip heatmap regression head: Predicts a Gaussian heatmap with peaks at ground truth tip locations.

3. Global placement classifier: Global average pooling followed by a fully connected layer produces 11-dimensional logits.

B2 cannot associate detected segments or tips with specific device instances, and its tip predictions carry no uncertainty estimate.

The central idea is breaking catheter detection into local fragment proposals. At each spatial location, the fragment proposal heads jointly predict:

1. Fragment heatmap ĥ(p) ∈ [0,1]: confidence that location p is a fragment center.

2. Sub-pixel offset ô(p) ∈ R2: refinement of the fragment center.

3. Geometry: endpoint offsets ê1(p), ê2(p) ∈ R2 and fragment length l̂(p) ∈ R+.

4. Tangent direction τ̂(p) ∈ R2: unit vector along the local catheter direction.

5. Discriminative embedding ê(p) ∈ R8: associative embedding vector trained to pull fragments from the same device together and push fragments from different devices apart.

6. Tip mean µ̂tip(p) ∈ R2: where this fragment thinks the device tip is.

7. Tip log-variance ŝ(p) ∈ R2: log-diagonal entries of the predicted tip covariance (heteroscedastic formulation following Kendall & Gal 2017), with a floor of 4 px2 per axis.

8. Fragment status probability q̂(p) ∈ [0,1]: probability that the fragment belongs to a foreground device.

An EdgeMLP scores candidate edges between fragment pairs by concatenating discriminative embeddings with geometric cues (relative position, tangent alignment, length ratio, tip prediction consistency). Each fragment considers at most 6 nearest spatial neighbors; edges scoring below 0.5 are pruned.

Device instances are recovered via constrained Union-Find, processing edges in decreasing order of score. Merges require:

1. Edge threshold ≥ 0.5.

2. Maximum degree of 2 neighbors within a cluster (enforcing linear, non-branching catheter topology).

3. Tip Mahalanobis consistency: Tip predictions must agree under predicted uncertainties (squared Mahalanobis distance below χ2 threshold of 9.21).

Each fragment contributes a tip estimate with predicted covariance. These are fused via iteratively-reweighted least squares (IRLS) with a Huber kernel over 3 iterations. Fragments with lower predicted uncertainty contribute more; the Huber kernel downweights outlier fragments. The fused covariance has a floor of 4 px2 per diagonal element.

A PlacementHead aggregates backbone features along the reconstructed path and at the fused tip location, producing placement logits for labels relevant to the device family. For weak supervision, a noisy-OR aggregation bridges per-device predictions to image-level labels: if any device triggers a positive label, the image-level prediction is positive.

Phase 1: Strong localization (35 epochs). Only strongly-annotated images are used. All fragment proposal heads, EdgeMLP, and segmentation/tip heads are trained with the full localization loss. AdamW optimizer with initial learning rate 3×10−4, weight decay 10−4, cosine annealing with 5-epoch linear warmup, gradient clipping at norm 5.0.

Phase 2: Hybrid fine-tuning (25 epochs). The full dataset enters training including weakly-labeled images. Learning rate drops to 7.5×10−5. PlacementHead and noisy-OR losses are activated.

RANZCR CLiP: 30,083 chest X-rays from 3,255 patients, spanning ETT (2,994 traced instances), NET (3,219 instances), CVC (11,629 instances), plus 157 Swan-Ganz catheters. 9,095 images (30.2%) carry device trace annotations totaling 17,999 traced instances. All images resized to 768 × 768. Patient-level 5-fold cross-validation (seed 20260730).

• Detection metrics: Hungarian algorithm matching; true positive requires both centerline fraction overlap and tip fraction distance ≤ 0.10. Reports sensitivity and precision.

• Tip localization error: Euclidean distance normalized by image diagonal (%diag).

• Placement classification: AUPRC for 5 primary placement labels (ETT-Abnormal, ETT-Borderline, NGT-Abnormal, NGT-Incompletely Imaged, CVC-Abnormal), reported as macro AUPRC.

• Association quality (B7 only): BCubed precision, recall, and F1.

• Uncertainty calibration (B7 only): 95% coverage metric—fraction of ground truth tips falling within predicted 95% confidence ellipse.

• Bootstrap confidence intervals: 2,000 patient-level bootstrap iterations with bias-corrected and accelerated (BCa) confidence intervals.

Metric B2 [95% CI] B7 [95% CI] Δ (B7−B2) [95% CI]

Tip error (%diag) ↓ 0.883 [0.832, 0.929] 1.725 [1.634, 1.811] +0.842 [0.755, 0.929]*

Sensitivity ↑ 0.674 [0.661, 0.686] 0.851 [0.843, 0.858] +0.177 [0.164, 0.190]*

Macro AUPRC ↑ 0.216 [0.205, 0.228] 0.226 [0.215, 0.239] +0.010 [0.002, 0.020]*

Precision ↑ 0.113 0.371 +0.258

False positives ↓ 18,798 4,771 −14,027

BCubed F1 ↑ — 0.775 —

Tip 95% coverage — 0.948 —

Key findings: B7 finds 26% more device instances (0.851 vs. 0.674 sensitivity) with 3.3× higher precision (0.371 vs. 0.113), slashing false positives from 18,798 to 4,771. Placement classification improves modestly (0.226 vs. 0.216 macro AUPRC), suggesting that the benefit of the compositional approach comes more from finding the right devices than from classifying each one better.

B7 matches or exceeds B2 on every label. Gains are largest for NGT-Abnormal and NGT-Incompletely Imaged—exactly where expected given B7's dramatically better NET detection. On ETT and CVC labels, where B2 already detects most devices, the gap narrows.

The aggregate tip error increase is explained by device family stratification:

• NET detection gap: B7 detects 554 NET tubes for which B2 produces zero matches. These carry a median tip error of 3.845 %diag—high in absolute terms because nasogastric tubes are long, thin, and frequently obscured by overlapping anatomy. This is a detection success masquerading as a localization failure.

• ETT head-to-head: On ETT devices matched by both models, the tip error gap shrinks from 0.842 %diag to just 0.275 (0.895 vs. 1.170 median). B7's mean drops below B2's (1.386 vs. 1.630) and its 95th percentile falls dramatically (3.359 vs. 5.735). B7's median is slightly worse but its tail is far better—it produces fewer of the catastrophic localization errors that actually matter clinically.

• Compositor dead end: A sweep of 12 compositor configurations (varying Huber threshold, IRLS iteration count, covariance floor) changed aggregate tip error by less than 0.03 %diag. Tip localization precision is determined at the fragment prediction stage; fusion merely aggregates what the fragments already know.

B7 achieves BCubed F1 of 0.775—roughly 77.5% of fragment-to-device assignments are correct. Performance degrades in crowded images where multiple devices run in parallel.

B7 achieves 95% coverage of 0.948—close to the ideal 0.950—indicating well-calibrated predicted covariance ellipses. Both models end up comparably calibrated at the image level for placement classification.

When devices overlap, B7 separates them into distinct instances through fragment association, whereas B2 collapses them into a single segmentation blob. Tip uncertainty ellipses widen appropriately for partially occluded or foreshortened devices. The fragment-level representation handles devices extending beyond the field of view.

"Aggregate tip error conflates localization precision with detection coverage. A model that finds more devices—including harder ones—will inevitably show higher aggregate error, even if it localizes each individual device just as well or better on matched instances."

From a clinical standpoint, detection sensitivity and precision matter more. A malpositioned catheter that goes undetected poses greater risk than a detected catheter whose tip is localized with slightly less precision.

The authors urge future benchmarks to report tip error conditioned on matched devices and stratified by family, Without that stratification, evaluation rewards models that detect less.

• Per-device placement assessment: the left subclavian CVC is borderline rather than something CVC-related is abnormal.

• Calibrated tip uncertainty: confidence ellipses grow for cases where the model genuinely struggles.

• Intermediate representations at each stage support clinical interpretability.

• Bottom-up architecture handles variable device counts without requiring a fixed maximum.

• NET tip localization remains the weakest link: median error of 3.845 %diag reflects inherently difficult imaging characteristics—thin, low-contrast, frequently overlapping with esophagus and mediastinal structures.

• BCubed F1 of 0.775 means roughly one in five fragment assignments is wrong, mostly in crowded images.

• Single dataset (RANZCR CLiP) from a single institution; generalization across imaging protocols, scanner manufacturers, and patient populations remains an open question.

At 2.27M parameters, UCompCXR is an order of magnitude smaller than EfficientNet or ResNet-50 backbones (20–25M parameters) common in medical imaging, and runs in a single forward pass—no Monte Carlo sampling, no multi-model cascade. This makes it practical for edge deployment on hospital workstations or embedded devices.

UCompCXR achieves substantially higher detection sensitivity (+26%) and precision (3.3×) than a strong multi-task baseline on the same lightweight backbone, while introducing per-device association (BCubed F1 = 0.775) and calibrated tip uncertainty (95% coverage = 0.948) that the baseline cannot provide. The analysis surfaces a methodological point: aggregate tip error penalizes models for finding harder-to-localize devices. On matched devices, the compositional model has fewer catastrophic failures. At 2.27M parameters and a single forward pass, UCompCXR is practical for clinical deployment. The clearest path to improvement runs through better fragment-level tip prediction for difficult device types—particularly nasogastric tubes—followed by scaling to additional device families and validation on external multi-institutional datasets.

Improvements to AI Systems:

1. Compositional instance detection with graph-based clustering: Replace monolithic segmentation or global classification heads with a fragment-proposal network that detects local device segments, then associates them into distinct device instances via constrained union-find clustering (enforcing linear topology and tip Mahalanobis consistency). This enables handling variable numbers of overlapping devices without fixed output limits.

2. Heteroscedastic tip localization with precision-weighted fusion: Implement per-fragment tip predictions that output both mean and log-variance (covariance), then fuse across fragments using iteratively-reweighted least squares with a Huber kernel. This yields calibrated uncertainty ellipses (95% coverage ≈ 0.95) and downweights outlier predictions, improving robustness on partially occluded devices.

3. Noisy-OR weak supervision bridging: Train placement classification end-to-end on weakly-labeled images by aggregating per-device predictions via noisy-OR to match image-level labels. This leverages large unannotated datasets while preserving instance-level granularity, improving detection sensitivity by 26% and precision by 3.3× versus strong baselines.

4. Discriminative embeddings for device association: Learn 8-dimensional associative embeddings per fragment, trained to pull same-device fragments together and push different-device fragments apart. Combined with geometric cues (tangent alignment, length ratio, tip consistency) in an edge-scoring MLP, this achieves BCubed F1 of 0.775 for fragment-to-device assignment.

5. Stratified evaluation and matched-device metrics: Report tip localization error separately for matched devices and by device family, avoiding the detection–localization conflation where finding harder devices (e.g., nasogastric tubes) inflates aggregate error. This guides model improvement toward genuinely weak cases rather than rewarding under-detection.

6. Lightweight multi-scale feature fusion: Use MobileNetV3-Small encoder with CompactFPN (depthwise separable convolutions, channel reduction) to produce unified stride-4 features. At 2.27M parameters in a single forward pass, this enables deployment on resource-constrained clinical hardware without Monte Carlo sampling or cascades.

What the Improved AI System Can Do:

• Detect and individually assess multiple overlapping catheters/tubes in a chest X-ray in one forward pass, outputting per-device paths, tip locations with calibrated confidence ellipses, and placement labels (normal/borderline/abnormal).

• Find 26% more devices than multi-task baselines while reducing false positives by 75%, particularly improving detection of thin, low-contrast nasogastric tubes that prior systems miss entirely.

• Provide clinically actionable per-device statements (e.g., left subclavian CVC is borderline) instead of image-level aggregates, with uncertainty that widens appropriately for occluded or foreshortened devices.

• Train effectively on mixed datasets where only 30% of images have device trace annotations, using weak image-level labels to boost performance on the remaining 70%.

• Operate in real-time on hospital workstations or embedded devices due to its small parameter count and single-pass inference, enabling bedside deployment for ICU chest X-ray triage.

Assessing catheter and tube placement on chest X-rays is safety-critical yet tedious and error-prone. Current deep learning methods either classify placement globally -- losing track of which device is where -- or segment all devices into a single mask, making per-device assessment impossible when catheters overlap. We introduce UCompCXR, a compositional framework that detects local catheter fragments, associates them into device instances via graph-based clustering, fuses per-fragment tip predictions through precision-weighted Gaussian estimation, and classifies placement per device. On the RANZCR CLiP dataset (30,083 images, 5-fold patient-level CV with bootstrap CIs), UCompCXR detects 26% more devices than a strong multi-task baseline sharing the same MobileNetV3 backbone, with 75% fewer false positives and well-calibrated tip uncertainty (95% coverage = 0.948). The aggregate tip error rises -- but only because the model finds devices the baseline misses entirely, especially nasogastric tubes. On matched devices, catastrophic localization failures drop substantially. At 2.27M parameters in a single forward pass, the model is deployable on resource-constrained clinical hardware.