Instance Segmentation of Catheters in Chest X-ray Images

Chest X-ray (CXR) is frequently employed in emergency departments and intensive care units to verify the proper placement of central lines and tubes and to rule out related complications. The automation of the X-ray reading process can be a valuable support tool for non-specialist technicians and minimize reporting delays due to non-availability of experts. While existing solutions for automated catheter segmentation and malposition detection show promising results, the disentanglement of individual catheters remains an open challenge, especially in complex cases where multiple devices appear superimposed in the X-ray projection. Moreover, conventional top-down instance segmentation methods are ineffective on such thin and long devices, that often extend through the entire image. To overcome these limitations, this study proposes a deep learning approach based on associative embeddings for catheter instance segmentation. A LaneNet architecture with a pretrained HRNetV2-W30 backbone was trained on 8877 CXRs containing up to 4 cardiac catheters, extracted from the RANZCR CliP dataset. The segmentation branch of LaneNet was trained to output a segmentation map with Dice loss, while the embeddings branch was guided by a discriminative loss to map each pixel to a 3d embedding. Embeddings predicted as foreground by the segmentation branch were extracted and clustered with mean shift, obtaining the final instance segmentation map. Semantic segmentation results were evaluated with IoU (0.60±0.01) and Dice coefficient (0.74±0.01). Instance segmentation results underwent AP and AR evaluation, averaged over 0.2:0.05:0.6 IoU thresholds, obtaining AP=0.73±0.01 and AR=0.81±0.01. This approach based on associative embeddings was able to effectively perform instance segmentation on chest radiographs with multiple catheters, providing promising options for automated radiology reporting, advanced visualizations and early identification of mispositioned and interfering devices.