ianpan/cxr-lines-and-tubes

CXR Lines and Tubes

Model description

This repository contains a research pipeline for detecting and classifying endotracheal tubes, nasogastric tubes, central venous catheters, and Swan-Ganz catheters on frontal chest radiographs, together with carina localization.

The release contains four model groups:

1. an EfficientNet-B1 encoder with separate U-Net decoders for device-agnostic and device-family line/distal-tip masks;
2. an EfficientNetV2-B0/DeepLabV3+ carina heatmap and coordinate model;
3. a five-fold EfficientNet-B1 multilabel classifier using RGB plus predicted device and carina probability maps; and
4. a lightweight FastViT four-family presence classifier for ETT, NGT, CVC, and Swan-Ganz screening.

The segmentation and competition-classification models are trained on RANZCR CLiP. The presence classifier additionally uses a non-overlapping subset of NIH ChestX-ray14 with thresholded teacher labels. The models are not medical devices and are not validated for patient-care decisions.

Intended use

Intended uses are research, reproducibility, model comparison, educational visualization, and development of human-reviewed clinical research tools.

Do not use the models as the sole basis for diagnosis, tube repositioning, triage, or any other patient-care decision. A qualified clinician must review the original radiograph and relevant clinical context.

Inputs and outputs

Device model

• Input: one grayscale chest radiograph, directly resized to 1024 x 1024.
• Agnostic outputs: line/tube union and distal-tip union probabilities.
• Family outputs: independent line and distal-tip probabilities for ETT, NGT, CVC, and Swan Ganz.
• Each device polyline contributes one distal-tip target: the most inferior endpoint for ETT/NGT and the endpoint closest to normalized image center for CVC/Swan-Ganz. An incompletely imaged NGT uses its most inferior visible endpoint, which is not necessarily its true anatomical tip.

Carina model

• Input: one grayscale chest radiograph, directly resized to 512 x 512.
• Outputs: carina heatmap, differentiable normalized coordinate, and peak coordinate.

Classifier

• Input: a 512 x 512 six-channel tensor consisting of RGB plus agnostic line, distal-tip, and carina probabilities.
• Output labels:
  • ETT Abnormal
  • ETT Borderline
  • ETT Normal
  • NGT Abnormal
  • NGT Borderline
  • NGT Incompletely Imaged
  • NGT Normal
  • CVC Abnormal
  • CVC Borderline
  • CVC Normal
  • Swan Ganz Catheter Present

The internal classifier also predicts a no-ETT class to form a four-way ETT softmax; that auxiliary output is excluded from the competition mean AUROC.

Lightweight device-presence classifier

• Input: one grayscale chest radiograph, directly resized to 384 x 384.
• Outputs: independent ETT, NGT, CVC, and Swan-Ganz presence probabilities.
• Training includes an inexpensive three-block depthwise-separable U-Net decoder for auxiliary family centerline supervision.
• Only the encoder, pooling layer, and four-logit classification head are needed for production screening.

Training data

The training set contains 30,083 RANZCR CLiP radiographs. Device polylines are available for 9,095 images, and 5,275 images have a valid single-point carina annotation under this project's filtering rule.

The presence-classifier training set first adds 17,869 non-RANZCR NIH radiographs whose CVC ensemble probability is below 0.95, a cutoff selected after manual review. A second controlled expansion adds 13,579 studies from previously unused patients only when all five FastViT teachers assign CVC probability below 0.001. There is at most one expansion study per patient. All 31,448 NIH rows are treated as CVC-negative, are training-only, and are excluded from every validation fold. ETT, NGT, and Swan-Ganz labels are thresholded at 0.5 from the original five-model B1 ensemble.

On exact concatenated hard-label RANZCR OOF predictions, the expansion improved CVC AUROC from 0.976654 to 0.977892 and mean family AUROC from 0.992299 to 0.992549. CVC improved in four of five folds. These NIH labels remain model-derived rather than verified ground truth.

The release links to the original datasets and does not redistribute source radiographs or annotations. Code is MIT-licensed. The model weights are released for non-commercial research, reproducibility, and educational use under the terms described in LICENSE.md.

Training procedure

• Five patient-disjoint folds with approximately balanced competition labels.
• Direct square resize without aspect-ratio preservation or padding.
• ImageNet initialization for all encoders/classifier backbones.
• Device Stage 1: distal-v4 B1 regular U-Net, 1024, 30 epochs, BCE plus low-weight line and distal-tip soft Dice.
• Device Stage 2: frozen eval-mode Stage 1 encoder, separate family U-Net decoder, 1024, 15 epochs; Swan-Ganz is sampled to approximately 10% prevalence.
• Carina: B0 DeepLabV3+, 512, 30 epochs, BCE plus coordinate Smooth L1.
• Classifier: B1, 512, 30 epochs, BF16 mixed precision.
• Segmentation features for annotated classifier images are strictly OOF; task-unannotated images use a five-fold probability ensemble.

Evaluation

Completed segmentation results:

Task	Metric	Five-fold mean
Device agnostic lines	Dice at 0.5	0.7501
Device agnostic distal tips	Dice at 0.5	0.6560
Device family	Dice	0.5855
Device family lines	Dice	0.5906
Device family distal tips	Dice	0.5804
Carina	Heatmap Dice	0.5236
Carina	Normalized landmark error	0.0145
Carina	PCK at 2% diagonal	0.7834

Reference five-channel classifier results:

Metric	Value
Raw OOF mean AUROC	0.96742
Fold-rank-normalized OOF mean AUROC	0.96769

Six-channel classifier best-checkpoint validation results:

Fold	AUROC
0	0.9652
1	0.9699
2	0.9697
3	0.9682
4	0.9690
Mean	0.9684

The six-channel mean averages independently selected fold checkpoints and is not a concatenated OOF AUROC.

Kaggle late-submission evaluation

Matched ensembles used the same number of segmentation teachers and classification folds. Model probabilities are averaged in probability space; no rank normalization is applied to the Kaggle test predictions. Scores are from late submissions to the original RANZCR CLiP leaderboard.

Ensemble	Notebook version	Private AUROC	Public AUROC
K1	4	0.96959	0.96085
K2	4	0.97143	0.96911
K3	1	0.97256	0.96967
K4	1	0.97273	0.97014
K5	1	0.97288	0.97059

K5 distal-v4 is the selected release ensemble. Scores improved monotonically with ensemble size on both leaderboard splits. Relative to the original two-endpoint K5, distal-v4 was 0.00026 lower privately and 0.00139 lower publicly, while providing the clinically useful single-distal-endpoint contract. These leaderboard scores evaluate the competition's 11-label classification task and do not establish clinical performance.

Potential future additions:

• Add a concatenated six-channel OOF AUROC if it is generated.
• Add per-label AUROC.
• Include confidence intervals or fold standard deviation where appropriate.

Limitations and biases

• Single competition dataset; external-site generalization is unknown.
• Synthetic thick polyline masks are not precise device-boundary annotations.
• The incompletely imaged NGT target is a visible inferior endpoint rather than a guaranteed anatomical tip.
• Rare Swan-Ganz examples limit family segmentation reliability.
• The presence classifier's NIH labels are model-derived and can contain label errors despite the manually selected CVC cutoff.
• The remaining NIH cohort has no hard device-presence labels and was selected by an earlier model. Its prediction distribution cannot establish prevalence or calibration.
• Devices can overlap, exit the field of view, or be incompletely imaged.
• Performance may change with acquisition hardware, post-processing, image orientation, patient population, anatomy, disease burden, and unseen devices.
• AUROC does not establish calibration or a safe clinical operating point.
• OOF/ensemble feature routing matches this research protocol and must be reproduced exactly for comparable results.

Ethical and safety considerations

The model can be confidently wrong. Visual overlays may encourage automation bias even when localization is inaccurate. Interfaces should display uncertainty, retain the original image, identify model/version provenance, and require human review. Do not suppress or replace the source radiograph with generated masks.

Hugging Face Space integration

The companion Space runs the lightweight four-family presence classifier first. When a configurable family threshold is reached, users may run the larger device segmentation, carina localization, and 11-label classification pipeline. The model repository contains every checkpoint from every fold rather than separate reduced ensembles. Models are loaded lazily, and users can choose the number of folds independently for presence screening, device segmentation, carina localization, and final classification.

Recommended profiles:

Profile	Presence folds	Device folds	Carina folds	Classifier folds	Intended use
screen	5	0	0	0	Reliable low-cost device-family screening
fast	1	1	1	1	Lowest-latency full result
balanced	5	3	3	3	Default speed/performance tradeoff
max	5	5	5	5	Highest-performing release ensemble

The distal-v4 K3 classifier pipeline reached 0.97256 private and 0.96967 public AUROC, close to K5 at 0.97288 and 0.97059. This supports K3 as the interactive default while retaining K5 for batch or maximum-quality inference.

Fold counts resolve to explicit, versioned fold-ID lists rather than depending on filesystem order. Advanced callers may provide exact fold IDs. The default K1–K5 device, carina, and classifier selections must reproduce the matched ensembles used for the Kaggle results; arbitrary unmatched combinations are supported but are not covered by those leaderboard measurements.

Presence outputs are raw sigmoid model scores with no post-hoc probability transformation. These scores must not be interpreted as clinical probabilities.

The companion Space uses 0.5 as an explicitly uncalibrated raw-score screening policy. If all four device-family scores are below 0.5, the larger pipeline is skipped unless the user selects the force-full-pipeline option. Changing the number or identity of folds changes the score distribution, so custom fold selections must not silently treat this threshold as calibrated.

The companion interface:

• display the original radiograph alongside continuous-probability overlays;
• report device-presence probabilities and the exact model version;
• display the active profile and fold IDs for every model component;
• allow independent fold-count overrides and a force-full-pipeline option;
• keep screening thresholds configurable until externally validated;
• explain when the larger pipeline was skipped; and
• present a prominent research-only, non-clinical-use warning.

The production runtime wrapper must return raw scores and identify any raw-score threshold used for staged execution. No internal calibration parameters are distributed with or loaded by the public release.

Release contents

• Five device-agnostic checkpoints.
• Five device-family decoder checkpoints.
• Five carina checkpoints.
• Five final classifier checkpoints.
• Final device-family presence classifier checkpoints.
• Versioned config module for every model group.
• Exact class/channel ordering.
• Preprocessing and normalization code.
• Ensemble profiles and inference runtime.
• Checkpoint SHA-256 hashes.
• Dependency requirements.
• Final Kaggle K1-K5 public/private metrics.
• Final CV metrics for the device-family presence classifier.
• Dataset, code, and model-weight licensing statement.
• Dataset citations and source repository.
• Space inference wrapper and UI.
• Versioned fold-selection manifest and screen/fast/balanced/max runtime profiles.

Reproducibility

The source repository records configs, losses, metrics, fold generation, OOF routing, and negative experiments. See docs/RANZCR_PIPELINE.md and docs/EXPERIMENT_HISTORY.md.

Source: https://github.com/i-pan/cxr-lines-and-tubes

Author and contact: Ian Pan, https://ianpan.me

Dataset citations

• Royal Australian and New Zealand College of Radiologists. RANZCR CLiP - Catheter and Line Position Challenge. Kaggle.
• Wang X, Peng Y, Lu L, Lu Z, Bagheri M, Summers RM. ChestX-ray8: Hospital-Scale Chest X-Ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases. CVPR 2017, 2097–2106.
• NIH ChestX-ray14 download: https://nihcc.app.box.com/v/ChestXray-NIHCC.

Space input and ETT–carina measurement

The companion Space accepts PNG, JPEG, and single-frame monochrome DICOM. DICOM images are transformed with the modality LUT and pydicom apply_voi_lut, converted to 8-bit, and inverted when the photometric interpretation is MONOCHROME1.

When the model predicts ETT presence and localizes both the distal ETT tip and carina, the Space reports Euclidean distance in the original upload geometry. It uses centimeters when DICOM pixel spacing is available and pixels otherwise. This automated measurement has not been clinically validated and is presented for research and demonstration only.