Deep Learning–Based Quantitative Assessment of Pneumothorax in Neonatal Chest Radiographs

Technical Summary: Deep Learning–Based Quantitative Assessment of Pneumothorax in Neonatal Chest Radiographs

Problem Statement
Pneumothorax is a life-threatening condition in neonates requiring rapid detection and accurate severity assessment to guide intervention. In Neonatal Intensive Care Units (NICUs), manual interpretation of supine chest radiographs is labor-intensive, time-consuming, and prone to inter-observer variability and diagnostic delays. While deep learning (DL) algorithms have demonstrated high performance in binary classification (detecting the presence of pneumothorax), existing systems largely fail to provide clinically relevant quantification of severity. Furthermore, current manual methods for estimating pneumothorax size involve complex formulas and lack standardized guidelines for pediatric patients. There is a specific gap in DL-based semantic segmentation applied to neonatal chest radiographs for objective, pixel-level quantification, particularly in cases involving extrathoracic extension.

Methodology
This multicenter retrospective study utilized 3,443 neonatal chest radiographs with pneumothorax collected from 11 university hospitals in Korea. The dataset was manually annotated and verified by three pediatric experts. The data was split at the patient level into training (n=2,754), validation (n=344), and test (n=345) sets. Additionally, thoracic segmentation masks for 60,931 images were generated via an AI-assisted pseudo-labeling pipeline and refined by experts.

• Model Architecture: The study employed two independent UNet++ models with ResNet-34 encoders. One model was trained to segment the pneumothorax region, and the other to segment the thoracic cavity. UNet++ was selected for its enhanced skip connections, which improve boundary continuity for thin pleural lines and small regions.
• Preprocessing: Images were converted to grayscale, resampled to 256×256 pixels, and center-cropped to 224×224 pixels. Standard data augmentation (brightness/contrast jitter, Gaussian blur, rotations, horizontal flips) was applied during training.
• Training: Models were trained using binary cross-entropy loss and the Adam optimizer (initial learning rate 1×10⁻⁴) for up to 200 epochs. Experiments were conducted on a Windows 11 workstation with four NVIDIA GeForce RTX 4090 GPUs using PyTorch 2.5.1.
• Quantification Strategy: Pneumothorax severity was quantified by calculating the pneumothorax-to-thorax area ratio derived from pixel-level masks. The system distinguished between intrathoracic (within boundaries) and extrathoracic (extending beyond boundaries) pneumothoraces. Severity was stratified into four ratio-based categories (<10%, 10–20%, 20–30%, ≥30%).
• Evaluation: Performance was measured using the Dice Similarity Coefficient (DSC) to assess spatial overlap between predicted and reference masks.

Key Results

• Segmentation Performance: In the test dataset (n=345), the pneumothorax segmentation model achieved a median DSC of 0.9088. The thorax segmentation model achieved a median DSC of 0.9386 on the test set.
• Severity-Dependent Accuracy: Segmentation accuracy improved as pneumothorax size increased. The median DSCs for pneumothorax-to-thorax ratios were:
  • <10%: 0.8366
  • 10–20%: 0.9191
  • 20–30%: 0.9452
  • ≥30%: 0.9742
• Extrathoracic Cases: The model successfully identified and segmented extrathoracic pneumothorax in 50 cases, maintaining a robust median DSC of 0.9531 across severity groups (ranging from 0.8840 to 0.9580).
• Qualitative Validation: Visual inspection of representative cases confirmed that automated segmentation closely matched manual ground truth for thorax, intrathoracic, and extrathoracic regions, with DSC values exceeding 0.96 in specific examples.

Key Contributions

1. Quantitative Assessment: The study introduces a DL-based system that moves beyond binary classification to provide an objective, reproducible quantitative measure of pneumothorax severity using pixel-level semantic segmentation.
2. Extrathoracic Capability: The system effectively handles complex cases where pneumothorax extends beyond the thoracic cavity, a scenario often challenging for standard detection algorithms.
3. Small Lesion Performance: Unlike previous studies that showed significant performance drops for small lesions, this model maintained stable performance (DSC 0.8366) even in cases with a pneumothorax-to-thorax ratio of less than 10%.
4. Clinical Workflow Integration: By automating the calculation of area ratios, the system aims to reduce the time and complexity associated with manual severity estimation in NICUs.

Significance and Claims
The authors claim that this deep learning-based system enables accurate detection, localization, and quantitative severity assessment of pneumothorax in neonatal chest radiographs. The study posits that this quantitative approach offers a reproducible method for severity stratification that may facilitate the earlier recognition of clinically significant pneumothorax and support timely, informed clinical decision-making.

The authors acknowledge limitations, including the use of a single established architecture (UNet++ with ResNet-34) and a limited number of extrathoracic cases due to their rarity. They emphasize that while technical validation relied on DSC, the clinical relevance was established through robust performance across stratified severity groups. The work concludes that the proposed algorithm provides a feasible tool for objective severity estimation, potentially enhancing management in neonatal care.