Open dataset and deep learning model for intelligent diagnosis of neonatal respiratory distress syndrome and aspiration syndrome in newborns.

Neonatal pulmonary diseases such as aspiration syndrome (AS) and respiratory distress syndrome (NRDS) require timely diagnosis, yet manual interpretation of neonatal X-rays is labor-intensive and subjective. To support AI-assisted diagnosis, we constructed the first Chinese neonatal pulmonary ailment dataset (NPA), covering both normal and diseased cases of varying severity. However, the NPA dataset exhibits severe class imbalance, and traditional augmentations randomly mix lesions, often corrupting pathological semantics. To address this, we propose a Polluted CutMix framework that selectively blends normal and diseased images, ensuring meaningful lesion synthesis, and an uncertainty-aware module that filters unreliable pseudo-labels during training. The novelty of this work lies in (1) introducing the first neonatal pulmonary dataset and (2) unifying targeted augmentation with uncertainty modeling for robust learning under data imbalance. On the NPA dataset, our approach surpasses existing baselines by 4-7% in classification accuracy, demonstrating improved generalization and diagnostic reliability. We hope that our novel framework and dataset will inspire further research in this field.