Automated Progressive Label Refinement for Nasogastric Tube Segmentation in Chest X-Rays

초록

This study addresses omission-dominant annotation noise in non-expert nasogastric (NG) tube labels by proposing a Progressive Label Refinement strategy guided by the Prediction-to-Annotation Ratio (PAR). PAR quantifies the directional discrepancy between the model-predicted tube extent and the current annotation, enabling selective replacement of likely under-annotated training labels without modifying the segmentation architecture. Using nnU-Net as the primary backbone, the proposed strategy improved the Dice similarity coefficient from 0.828 to 0.844 on the internal validation set, from 0.8124 to 0.8306 on the internal test set, and from 0.808 to 0.816 on the external MIMIC-CXR test set. Robustness experiments with synthetic under-segmentation noise showed that intermediate refinement thresholds consistently provided the best balance between correction and stability, with the strongest results typically observed around $au=0.5$ . Additional validation with DeepLabV3+ also yielded an improvement from 0.8114 to 0.8231 on the internal test set, providing preliminary evidence that the refinement logic can transfer across two backbones. However, prolonged refinement caused late-stage performance collapse in the supplementary experiment, indicating that explicit stopping and stabilization mechanisms are necessary. These findings show that PAR-guided label refinement is a practical data-centric strategy for improving NG tube segmentation when expert-quality annotations are limited.

키워드