Abstract
Accurate lung field segmentation in chest radiographs is essential for monitoring cardiac health in hemodialysis patients. This paper presents a novel deformable model that integrates general population shape statistics with patient-specific historical data. By dynamically transitioning from a population-based model to one informed by an individual's unique anatomy over a series of scans, the system achieves superior accuracy and temporal stability compared to traditional methods.
Introduction
For patients undergoing hemodialysis, maintaining "dry weight"—the ideal body weight without excess fluid—is critical for preventing cardiovascular complications. Clinicians often monitor fluid status by measuring heart size in chest X-rays as a proxy. However, precise heart measurement requires accurate segmentation of the surrounding lung fields. This task is complicated by anatomical variations and image artifacts, such as rib shadows, which can mislead standard edge-detection algorithms. While population-level models provide a baseline, they often fail to capture the specific anatomical nuances of individual patients over time.
Methods
The proposed approach utilizes a deformable model guided by a multi-component energy function. Key technical innovations include:
- Modified SIFT Descriptors: To overcome the distraction of rib edges, the model employs Scale-Invariant Feature Transform (SIFT) descriptors. These capture complex local textures, allowing for more robust identification of lung boundaries.
- Hybrid Shape Statistics: The model's shape is constrained by a combination of general population data and patient-specific historical imaging.
- Dynamic Weighting and Iterative Refinement: A weighting parameter balances these two data sources. Initially, the model relies on population data; after approximately five scans, it shifts focus toward the patient’s own historical statistics. Furthermore, an iterative process uses new scan data to refine previous segmentations in the series.
Results
The algorithm was evaluated using a dataset of serial frontal chest radiographs from 39 patients, many with up to 17 monthly scans. Performance was compared against Active Shape Models (ASM) and traditional Snake models.
The proposed method achieved a 93.8% ± 0.58% overlap with expert manual segmentation, significantly outperforming ASM (92.6%) and Snake models (78.7%). The average contour distance error was reduced to 1.76 pixels. Crucially, the method demonstrated high temporal stability, maintaining consistent segmentation across long-term series despite variations in image quality or patient pose.
Discussion and Conclusion
The integration of patient-specific statistics represents a significant advancement in medical image segmentation. By "learning" an individual's anatomy, the model moves beyond the limitations of "one-size-fits-all" population models. While the system requires a "warm-up" period of several scans to reach peak performance, its robustness makes it ideal for chronic disease monitoring. Future work may extend this framework to other serial imaging tasks, such as tracking tumor progression or chronic pulmonary conditions.
Summary updated