TL;DR
This study reveals that current state-of-the-art methods for segmenting lung regions on chest X-rays (CXRs) significantly underestimate the actual lung volume, missing substantial portions. This underestimation, driven by the 2D projection of 3D anatomy, could lead to missed diagnoses and highlights the need for more comprehensive segmentation approaches, potentially integrating 3D imaging data.
What Problem Does This Paper Address?
Chest X-rays (CXRs) are a cornerstone of medical imaging, but their interpretation, especially for automated analysis, relies heavily on accurate segmentation of anatomical structures like the lungs. The challenge lies in the fact that CXRs are 2D projections of complex 3D anatomy. This means that parts of the lungs can be obscured by other structures such as the heart, mediastinum, or diaphragm. Current methods for creating lung segmentation masks on CXRs, often referred to as the "state-of-the-art," typically follow the contours of these adjacent organs. This approach, while standard, may not capture the full extent of the lung fields, leading to an underestimation of lung volume. This underestimation is problematic because it could lead to missed detection of abnormalities located in these obscured regions, potentially resulting in clinical errors. The paper aims to quantify this underestimation by comparing 2D CXR-based lung masks with 3D lung volumes derived from CT scans.
Methods
The researchers employed a retrospective study design using low-dose computed tomography (LDCT) scans from 55 healthy individuals. To simulate CXR images, they generated synthetic X-ray images (Digitally Reconstructed Radiographs or DRRs) by projecting the 3D CT data onto a 2D plane.
Two experienced radiologists then created two types of lung masks:
- 3D CT masks: These were generated semi-automatically from the 3D CT data using a specialized lung segmentation module. These served as the "ground truth" for total lung volume.
- 2D DRR masks: These were manually created from the synthetic X-ray (DRR) images, following the current state-of-the-art method of outlining adjacent anatomical structures.
The core of their analysis involved comparing these two types of masks. They calculated the volume of overlapping regions (where both masks agreed) and non-overlapping regions (where the 3D CT mask extended beyond the 2D DRR mask). The volume of these non-overlapping regions, representing the "obscured lung regions," was then expressed as a fraction of the total lung volume derived from the 3D CT masks. Statistical analyses, including Dice Score Coefficient (DSC) and Jaccard Index (JI) for spatial agreement, and dependent samples t-tests for volume comparisons, were used to evaluate the results.
Datasets
The study utilized a dataset of 55 anonymized, non-contrast chest low-dose CT (LDCT) examinations. These scans were part of a national screening program conducted between 2010 and 2018, involving 60 medical centers. The original dataset contained 37,839 examinations, from which the 55 participants were selected based on specific inclusion and exclusion criteria (e.g., age between 50-60, no history of lung disease or smoking, and no major findings on the original radiology report).
Key Insights and Ideas
- The "state-of-the-art" is significantly underestimating lung volume: The most striking finding is that current methods for creating lung masks on CXRs miss a substantial portion of the lung volume, averaging around 27.3% for both lungs.
- 2D projection is the root cause: The inherent limitation of CXRs being 2D projections means that anatomical structures can overlap, leading to portions of the lungs being hidden from view and thus excluded by masks that follow external contours.
- Subjectivity in manual segmentation: The paper points out that current lung mask creation lacks well-defined criteria, leading to high subjectivity among annotators and variability in how masks are generated.
- 3D imaging offers a solution: Integrating 3D imaging modalities like CT and MRI holds significant potential for creating more precise and comprehensive 2D lung masks by providing a complete anatomical context.
Key Results
- Significant underestimation: The mean volume fractions of lung regions not covered by synthetic X-ray masks were 22.8% for the right lung, 32.9% for the left lung, and 27.3% for both lungs. These figures were consistent between the two radiologists.
- Excellent spatial agreement for masks: Despite the underestimation, the spatial agreement between the masks created by the two radiologists was excellent, with high DSC and JI scores for both 3D CT masks and 2D DRR masks.
- Volume differences are significant: The total lung volume calculated from CT masks and the lung volume calculated from DRR masks differed significantly in all cases.
Limitations and Future Work
The authors acknowledge several limitations:
- Use of DRRs instead of original CXRs: While DRRs are a well-established method for simulating CXRs, they are not identical to actual CXR images.
- Examination position discrepancies: CT scans are typically performed with patients in a supine position, while CXRs can be obtained in standing or supine positions. This difference in patient posture could lead to discrepancies between DRRs and actual CXRs.
- Homogeneous, healthy patient group: The study focused on healthy individuals, and the results might vary in patients with lung abnormalities (e.g., pneumonia, pleural fluid, cardiomegaly) that could affect lung volume or chest organ positions.
Future work suggested by the authors includes:
- Developing standardized criteria: Establishing clear guidelines for manual lung mask creation to reduce subjectivity and improve consistency.
- Enhancing automatic segmentation: Improving algorithms to better capture the entire lung volume, potentially by training on datasets that include obscured regions.
- Integrating 3D data: Leveraging 3D imaging modalities like CT and MRI to create more accurate and comprehensive 2D lung masks.
- Proposing new mask creation strategies: The paper suggests three potential solutions: creating a single unified mask that includes obscured regions, creating separate masks for visible and obscured regions, or defining a standardized chest region encompassing the lungs.
Summary updated