NCT-CXR: Enhancing Pulmonary Abnormalities Segmentation on Chest X-ray using Improved Coordinate Geometric Transformation

Medical image segmentation, especially in chest X-ray (CXR) analysis, encounters substantial problems such as class imbalance, annotation inconsistencies, and the necessity for accurate pathological region identification. This research introduces NCT-CXR, a robust framework that enhances semantic segmentation in CXR images using an improved coordinate-geometric transformation strategy. NCT-CXR integrates carefully calibrated geometric transformations with intensity-based augmentations, ensuring spatial accuracy throughout the augmentation process. The framework was evaluated on the NIH Chest X-ray dataset comprising 1,061 images across nine pathological categories. NCT-CXR has four different coordinate transformation models, i.e. discrete rotations at (-10°, +10°), discrete rotations at (-5°, +5°), and mixed rotation augmentation. Semantic segmentation was performed using YOLOv8 with optimized hyperparameters. Non-parametric statistical analysis using Kruskal-Wallis test revealed significant differences in precision metrics (H= 14.874, p = 0.001927), while other performance metrics remained stable. Subsequent Nemenyi post-hoc analysis demonstrated that discrete-angle rotations at (-5°, +5°) and (-10°, +10°) significantly outperformed mixed rotations (p = 0.013806 and p = 0.005602 respectively). These models achieved particularly high precision in pneumothorax detection (0.829 and 0.804 respectively), emphasizing the effectiveness of controlled geometric transformations for conditions with well-defined anatomical boundaries. These findings demonstrate the efficacy of NCT-CXR in producing clinically relevant segmentation outcomes and underscore the importance of augmentation design in pathology-specific model performance. Future work will explore the generalizability of this approach across diverse imaging modalities and its applicability to a broader spectrum of thoracic conditions.