NCT-CXR: Enhancing Pulmonary Abnormality Segmentation on Chest X-Rays Using Improved Coordinate Geometric...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computer Vision

Background:

• Medical image segmentation, particularly for chest X-rays (CXRs), faces challenges like class imbalance and inconsistent annotations.
• Accurate identification of pathological regions in CXRs is crucial for diagnosis and treatment planning.
• Existing methods often struggle with the nuances of anatomical variations and subtle abnormalities.

Purpose of the Study:

• To develop NCT-CXR, a novel framework for precise and clinically reliable pulmonary abnormality segmentation in CXRs.
• To improve deep learning model performance by combining anatomically constrained data augmentation with expert-guided annotation refinement.
• To address label noise in large CXR datasets and enhance segmentation accuracy for thoracic conditions.

Main Methods:

• Implemented NCT-CXR framework utilizing discrete-angle rotations (±5°, ±10°) and intensity-based augmentations for data enrichment.
• Employed a clinically validated annotation refinement pipeline with OncoDocAI for generating high-quality, multi-label pixel-level segmentation masks.
• Selected YOLOv8 as the segmentation backbone for its efficiency, speed, and spatial accuracy.
• Utilized NIH Chest X-ray dataset for training and evaluation.

Main Results:

• NCT-CXR significantly improved segmentation precision, notably for pneumothorax (achieving 0.829 and 0.804 Dice scores with ±5° and ±10° rotations, respectively).
• Statistical analysis (Kruskal-Wallis, Nemenyi) confirmed the superiority of discrete-angle augmentation over mixed strategies (p < 0.014).
• The refined annotations and augmentation strategy led to more robust and accurate segmentation masks for nine thoracic conditions.

Conclusions:

• Clinically constrained data augmentation and high-quality annotation are vital for robust medical image segmentation models.
• NCT-CXR provides a practical and high-performance solution for integrating deep learning into radiological workflows.
• The framework demonstrates potential for enhancing diagnostic accuracy and efficiency in CXR analysis.