Automated Detection and Type Classification of Central Venous Catheters in Chest X-Rays

Central venous catheters (CVCs) are commonly used in critical care settings for monitoring body functions and administering medications. They are often described in radiology reports by referring to their presence, identity and placement. In this paper, we address the problem of automatic detection of their presence and identity through automated segmentation using deep learning networks and classification based on their intersection with previously learned shape priors from clinician annotations of CVCs. The results not only outperform existing methods of catheter detection achieving 85.2% accuracy at 91.6% precision, but also enable high precision (95.2%) classification of catheter types on a large dataset of over 10,000 chest X-rays, presenting a robust and practical solution to this problem.

💡 Research Summary

This paper addresses the clinically important task of automatically detecting central venous catheters (CVCs) and identifying their specific type (PICC, internal jugular, subclavian, or Swan‑Ganz) on routine anterior‑posterior chest X‑ray (CXR) images. The authors propose a hybrid pipeline that combines deep learning–based segmentation with handcrafted, anatomy‑aware features fed into a conventional random‑forest classifier.

Segmentation stage
A modified U‑Net architecture is trained to produce pixel‑wise probability maps of any CVC regardless of type. Because catheters occupy less than 1 % of the image area, the authors replace the standard Dice‑plus‑cross‑entropy loss with the exponential logarithmic loss, which better handles severe class imbalance. Training uses 1 500 AP CXRs from the NIH dataset that were manually annotated at the pixel level (608 images containing 359 IJ, 78 subclavian, 277 PICC, and 32 Swan‑Ganz catheters). The network converges with a Dice score around 0.78; qualitative results show that while many catheter fragments are correctly highlighted, some portions are missed and occasional false positives appear.

Feature extraction
From the U‑Net output the authors construct three groups of features:

1. Spatial priors – For each catheter class, clinicians’ contour annotations are averaged pixel‑wise, producing a probabilistic “shape prior” map. The segmentation output is multiplied element‑wise with each prior, and the resulting overlap is summarized by intensity histograms and Histogram‑of‑Oriented‑Gradients (HoG). This quantifies how well a candidate region matches the expected trajectory of a given catheter type.

2. Anatomical context – Separate U‑Nets segment key thoracic structures (clavicles, lungs, heart, mediastinum). Euclidean distance distributions between the catheter segmentation and the centroids of these structures are computed, providing contextual cues that help differentiate, for example, PICC lines (which run along the lateral chest wall) from subclavian lines (which pass near the clavicle).

3. Size and shape descriptors – Area, length, width of the segmented region, plus HoG of its contour, capture the overall geometry of the catheter.

These features are concatenated into a single vector for each image.

Classification stage
Two random‑forest (RF) models are trained. The first RF performs binary classification (catheter present vs. absent). The second RF outputs a multi‑label prediction for the four catheter types. Hyper‑parameters (number of trees, depth) are tuned via 5‑fold cross‑validation.

Datasets and experimental protocol

• Presence detection: 3 000 CXRs labeled for any external medical device (2 381 positive, 619 negative).
• Type classification: 10 746 CXRs with at least one catheter, labeled via a semi‑structured radiology report template and NLP clustering (4 249 PICC, 1 651 IJ, 201 subclavian, 192 Swan‑Ganz, plus other devices).

All experiments use a 60‑20‑20 split for training, validation, and testing, repeated across five folds.

Results

• Presence detection: The full feature set (spatial prior + size + anatomical context) yields 85.2 % accuracy, 91.6 % precision, 89.6 % recall, and an AUC of 79.4 %. Simpler baselines (VGG‑16, DenseNet) trained end‑to‑end on the raw CXRs achieve only ~20‑30 % accuracy, confirming that standard CNNs cannot reliably learn from such thin structures. Adding the segmentation mask as an extra channel improves AUC modestly but does not affect precision/recall.

• Type classification: The same feature‑driven RF reaches an average accuracy of 78.2 % and a weighted precision of 95.2 %. Performance is highest for PICC lines (recall > 80 %) because they dominate the dataset and have a relatively simple trajectory. Recall for the other three classes remains below 50 %, reflecting class imbalance and more complex shapes. All deep‑learning‑only approaches (VGG, DenseNet, or their feature‑extracted variants) fall far short, with accuracies around 30‑40 % and AUC near 50 %.

Discussion
The key contribution is the demonstration that integrating clinician‑derived spatial priors and anatomical context into a lightweight machine‑learning classifier can dramatically outperform pure deep‑learning pipelines for detecting and classifying tiny, elongated medical devices. The approach is data‑efficient (only 1 500 pixel‑level annotations needed) and robust to the poor signal‑to‑noise ratio inherent in CVC visualization. Limitations include reliance on priors derived from a single institution’s imaging protocol, which may limit generalizability, and the pronounced class imbalance that hampers recall for non‑PICC catheters. Future work should explore multi‑center validation, domain adaptation, synthetic data augmentation, and possibly integrating transformer‑based encoders to capture longer‑range dependencies while still leveraging the handcrafted priors.

Conclusion
The authors present a practical, high‑precision system for automated CVC detection and type classification on chest X‑rays. By coupling a U‑Net segmentation front‑end with spatial‑prior‑driven features and a random‑forest back‑end, they achieve 85 %+ accuracy for presence detection and 95 % precision for type classification, substantially surpassing existing deep‑learning baselines. This work illustrates a promising direction for combining expert knowledge with machine learning to solve challenging medical imaging problems where target structures are small, sparse, and highly variable.