Abstract
This paper proposes a structure-by-structure extraction method of spinal columns, ribs, intervertebral disks and spinal processes in abdominal X-ray CT images. Bone regions are extracted using binarization, and then spinal columns and ribs are obtained from the bone regions using morphological operators. The regions of intervertebral disks are emphasized. The positions of intervertebral disks are estimated using plane models, and spinal columns are divided into vertebra columns based on estimated positions. Especially, the 12th thoracic vertebra is identified based on the size of the connected ribs. They can be used as anatomical landmarks. The upper thoracic vertebrae and lumbar vertebrae are identified as well. Plane models can divide spinal columns, but cannot divide spinal processes correctly. Therefore, curved surface models are introduced in this paper. The proposed method was applied to actual CT images, and experimental results are shown.
Introduction
Interpreting abdominal CT scans is a high-stakes, labor-intensive task for radiologists and surgeons. The "Images 1" section of the Journal of the Japan Society of Computer Aided Surgery (2016) presents a collection of advancements in automated abdominal image analysis. These reports address the difficulty of automatically identifying and segmenting complex anatomical structures such as the spinal column, biliary tract, and fine vascular trees, which often suffer from low contrast and high noise in traditional CT imaging.
Methods and Feature Engineering
The researchers employ a diverse array of classical and modern computer vision techniques to move beyond simple intensity thresholding. Key methodologies include:
- Geometric and Curved Surface Modeling: Used to divide the spinal column into individual vertebrae and accurately segment spinal processes that standard plane models cannot handle.
- Structured Random Forest (SRF): Applied to 3D CT volumes for lymph node detection, utilizing "label patches" to capture local image structure more effectively than standard voxel-wise classification.
- Advanced Feature Selection: To distinguish target tissues from surrounding fat and noise, researchers incorporated Hessian-based blob/line enhancement filters, Local Binary Patterns (LBP), and positional features relative to anatomical landmarks like the liver center.
Key Results
The integration of these novel features and models has led to significant improvements in segmentation accuracy:
- Vascular Labeling: An automated tree construction method achieved 91.7% accuracy in classifying abdominal arteries across 50 clinical cases.
- Biliary Segmentation: The addition of positional relationship features improved the F-measure from 0.50 to 0.56, successfully isolating the bile duct from similar-looking tissues in Dual-Energy CT (DECT).
- Lymph Node Detection: The SRF-based approach reached a detection rate of 57.8% overall, with a higher performance of 71.4% for nodes smaller than 10mm.
Discussion and Future Work
While these methods show great promise for clinical surgical planning, challenges remain. Many techniques still struggle with "over-segmentation," mistaking fat or adjacent organs for the target structure. Furthermore, the current reliance on small-scale clinical datasets (n=5 to n=50) necessitates broader validation. Future work will focus on automating parameter selection to reduce the need for manual weight adjustment, making these computer-aided diagnosis (CAD) tools more accessible for routine clinical use.
Summary updated