UCompCXR catches 26% more catheters in chest X-rays with 75% fewer false positives

Assessing catheter and tube placement on chest X-rays is critical for patient safety, but current deep learning models 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. To fix this, Harshil Lodhiya introduced UCompCXR, a compositional framework that works in three stages: it detects local catheter fragments, groups them into device instances via graph-based clustering, and fuses per-fragment tip predictions using precision-weighted Gaussian estimation. This allows the model to identify each catheter as a separate object and classify its placement individually, even in complex overlapping cases.

Validated on the RANZCR CLiP dataset with 30,083 images using 5-fold patient-level cross-validation, UCompCXR outperformed a strong multi-task baseline sharing the same MobileNetV3 backbone. The model detected 26% more devices overall and reduced false positives by 75%. Its tip uncertainty estimates are well-calibrated, achieving 95% coverage at 0.948. Interestingly, the aggregate tip error increased—but only because UCompCXR finds devices (especially nasogastric tubes) that the baseline misses entirely; on matched devices, catastrophic localization failures dropped substantially. With just 2.27M parameters and a single forward pass, UCompCXR is deployable on resource-constrained clinical hardware, making it practical for real-world radiology workflows where speed and accuracy are crucial.