Abstract
Introduction
In this work, we present the first real-time fully automatic system for reconstruction of patient-specific 3D knee bones models using ultrasound raw RF data. The system was experimented on two cadaveric knees, and reconstruction accuracy of 2 mm was achieved.
Methods
To use the highest available contrast and spatial resolution in the ultrasound data, the raw RF signals were used directly to automatically extract the bone contours from the ultrasound scans. Figure 1 shows a sample ultrasound B-mode image for cadaver's distal femur, showing some of the scan lines raw RF signals as well as the final extracted contour using our method.
An ultrasound machine (SonixRP, Ultrasonix Inc) was used to scan the knee joint and the RF data of the scans are acquired by custom-built (using Visual C++) software running on the ultrasound machine. An optical tracker (Polaris Spectra, Northern Digital Inc) was attached to the ultrasound probe to track its motion while being used in scanning.
The scanning of the knee was performed at two flexion angles (full extension, and deep knee bend). At each position, the knee was fixed in order to collect scans that represent a partial surface of the bone (which will be later mutually registered to represent the whole bone's surface). Figure 4 shows fluoroscopy images of a patient's knee, showing the different articulating surfaces of the knee bones visible to the ultrasound at different flexion angles. Figure 5 shows a dissected cadaver's knee showing the articulating surfaces visible to ultrasound at 90 degrees flexion.
The custom-built software collects the RF data synchronized with the probe tracking data for each ultrasound frame. Each frame of the RF data is then processed to extract the bone contour. The bone contours are automatically extracted from the RF data frame with frame rate of 25 frames per second. Figure 2 shows a flowchart for the contour extraction process.
The extracted bone contours were then used by the our software, along with the ultrasound probe's tracking data, to reconstruct point clouds representing the bones' surfaces. These point clouds were then aligned to the mean model of the bone's atlas using ICP and integrated together to form 3D point cloud of the bone's surface. A 3D model of the bone is then reconstructed by morphing the mean model to match the point cloud. Figure 3 shows a flowchart for the point cloud and 3D model reconstruction process.
Results
The developed system was tested on two cadavers' knees. The cadavers' knees were CT-scanned and manually segmented. The reconstructed models using ultrasound were then compared to the segmented models. An average error of 2 mm was achieved. Figure 6 shows sample ultrasound RF signals, and their processed version and the extracted bone echoes. Figure 7 shows sample ultrasound frames and the extracted bone contours from them. Figure 8 shows the reconstructed point clouds and 3D models for two distal femurs and a proximal tibia.
