Structure-Aware Deep Segmentation of Nasogastric Tubes: A Benchmark of Modern Topological Losses and Augmentations

Accurate localization of nasogastric tubes (NGTs) in chest X-rays is essential for safe clinical placement. Yet, standard deep learning models relying on overlap-based loss functions often neglect topological continuity, resulting in fragmented segmentations. We present a systematic benchmark of topologyaware methods-including advanced loss functions and the CoLeTra data augmentation-within the nnU-Net framework. Our best-performing configuration, combining Skeleton Recall Loss with CoLeTra augmentation, yielded consistent improvements across internal and external datasets. Notably, in challenging cases with overlapping chest tubes, it achieved relative gains of 10.9% in Dice Similarity Coefficient (DSC) and 10.4% in centerline Dice (clDice) compared to the baseline. These results demonstrate the effectiveness of topology-preserving learning for segmenting thin anatomical structures under real-world clinical conditions, without introducing additional inference overhead.