Abstract
Accurate delineation of organs at risk (OARs) is critical in radiation therapy planning to maximize tumor dosage while minimizing damage to healthy tissue. This paper presents a semi-automatic segmentation method that combines a "standard man" atlas with active contour models. By utilizing polynomial warping for global alignment followed by "snake" refinement for local edge detection, the proposed approach aims to improve the speed and consistency of organ segmentation compared to traditional manual methods.
Introduction
In radiation therapy (RT), precision is paramount. Clinicians must accurately map OARs to ensure patient safety. Historically, this process has been manual and labor-intensive, requiring radiologists to outline organs slice-by-slice. This manual approach is not only time-consuming but also suffers from significant inter-observer variability.
The primary challenge in automating this process is the vast anatomical diversity among patients. A static template cannot account for individual geometric variations. This study addresses this by bridging the gap between a reference anatomical atlas and patient-specific CT data using a two-stage deformation and refinement strategy.
Methodology
The proposed semi-automatic segmentation procedure involves two primary stages:
- Global Alignment via Polynomial Warping: The Visible Human Male Computed Tomography (VHMCT) dataset serves as the reference "standard man" geometry. Target organs, such as the liver, spinal cord, and right lung, are manually outlined on this reference. To map these outlines to a new patient, a user identifies seven matching landmarks on both the reference and the patient's CT slice. A polynomial warping algorithm then transforms the reference outline to match the patient's general proportions.
- Local Refinement via Active Contour Models: To account for local anatomical nuances, "snakes" (active contour models) are employed. These mathematical splines are attracted to image gradients, allowing them to "shrink-wrap" around the actual organ boundaries. The polynomial warping provides a crucial "initial guess," placing the snake close enough to the target to avoid local minima and ensure successful convergence.
The study evaluated this method using 12 patient studies, with five CT slices per study matched to corresponding VHMCT slices.
Results and Discussion
The semi-automatic outlines were compared against manual delineations by a radiologist. The results demonstrated strong agreement for high-contrast organs:
- Liver: Area agreement ranged from -5% to +7%, with volume agreement within ±4%.
- Right Lung: Area agreement was between -11% and -2%, with volume agreement between -8% and -5%.
- Spinal Cord: While showing higher percentage variance in area (±28%) and volume (±21%), the absolute difference was minimal (±1cm²) due to the organ's small size.
The study highlighted that segmentation success is heavily dependent on the "attenuation difference" (contrast) between the organ and its surroundings. The liver proved most challenging due to its proximity to tissues with similar CT densities, occasionally causing the snake to "leak" into adjacent structures. Conversely, the high contrast of the spinal cord against surrounding vertebrae provided surprising reliability despite its small dimensions. Furthermore, the automated method produced more consistent CT-number values within the outlines, suggesting a reduction in errors related to the partial volume effect.
Conclusion
This research demonstrates the feasibility of using an atlas-based warping technique combined with active contours for OAR delineation. While currently limited to 2D slices and requiring manual landmarking, the method offers a path toward more consistent and efficient RT planning. Future work will focus on expanding the system to 3D warping, developing more sophisticated energy functions for low-contrast boundaries like the liver, and automating landmark detection to achieve a fully automated workflow.
Summary updated