TL;DR
This paper introduces a novel, fully automated method for segmenting mammary gland regions in non-contrast torso CT scans. By combining a probabilistic atlas of mammary gland and muscle locations with CT number distribution analysis, the approach achieves robust segmentation, paving the way for improved breast tumor detection in CAD systems.
What Problem Does This Paper Address?
Accurate identification of mammary gland regions is a crucial first step for breast tumor diagnosis using medical imaging. While digital mammography is common, it's a 2D projection and can miss 3D structural details. Volumetric CT scans offer a 3D view of the entire torso, including mammary glands, but segmenting these regions is challenging. Mammary glands lack a consistent CT number (density) and their location varies significantly between individuals and with age, making traditional segmentation methods (like simple thresholding or region growing) ineffective. This paper aims to solve this problem by developing an automated segmentation technique for these challenging regions.
Methods
The proposed method employs a two-pronged approach:
- Probabilistic Atlas Construction: The core of the method is a "probabilistic atlas." This atlas is built from a learning set of CT scans where mammary gland regions have been manually segmented by a radiologist. By normalizing and deforming these scans into a standard coordinate space, the researchers create probability maps indicating the likelihood of a voxel belonging to the mammary gland or chest muscle. This atlas captures the typical anatomical locations and variations.
- Mammary Gland Segmentation: For a new CT scan, the probabilistic atlas is deformed to match the specific patient's anatomy. This allows for the estimation of mammary gland and muscle locations. The CT number distribution of the mammary gland is approximated using a Gaussian function. Finally, a thresholding process, guided by the probability maps and density estimation, is used to segment the mammary gland regions, followed by a binary connected-component processing refinement. Nipple locations are also extracted as important landmarks to aid in the process.
Datasets
The study utilized non-contrast CT scans from 66 female patients, aged 40-80 years. The "gold standard" for evaluating the segmentation accuracy was generated by manual segmentation performed by a radiologist specialized in breast cancer. The paper mentions a "learning sample dataset" used for constructing the probabilistic atlas, but specific details on its size or composition beyond the 66 cases are not explicitly detailed.
Key Insights and Ideas
- Probabilistic Atlas for Anatomical Variability: The key innovation is the use of a probabilistic atlas to handle the inherent anatomical variability of mammary glands. Instead of relying on fixed shapes or densities, the atlas captures the likelihood of gland presence across a population.
- Combining Location and Density: The method effectively combines spatial information (from the probabilistic atlas) with intensity information (CT number distribution) to achieve more accurate segmentation than either approach alone.
- Nipple Location as a Landmark: Utilizing nipple locations as anatomical landmarks provides crucial anchors for aligning the probabilistic atlas and refining the segmentation.
- Addressing Over-extraction: The discussion highlights that over-extraction (false positives), particularly misclassifying chest muscle as mammary gland, is a significant challenge. This suggests that distinguishing between these similar-density tissues remains an area for improvement.
Key Results
The proposed automated segmentation scheme was evaluated on 66 patient cases. The primary metric used was the coincidence rate between the automated segmentation and the radiologist-defined gold standard.
- The mean coincidence rate was 0.82, with a standard deviation of 0.09.
- The authors state that results with a coincidence ratio larger than 0.8 were considered acceptable.
- Qualitative results shown in Figures 6 and 7 demonstrate visually plausible segmentation of mammary gland regions, though some cases (like those in Fig. 7) exhibit over-extraction.
Limitations and Future Work
- Over-extraction (False Positives): The authors acknowledge that over-extraction, specifically misclassifying chest muscle as mammary gland, was observed in some cases. This is attributed to the similar density between these tissues and the relatively small volume of mammary glands.
- Future Work: While not explicitly detailed, the discussion on over-extraction implies that future work could focus on improving the discrimination between mammary glands and surrounding muscle tissues. The paper suggests the method is useful as an initial stage for breast tumor detection in CAD systems, hinting at further integration into diagnostic workflows.
Summary updated