Automatic detection of lumbar anatomy in ultrasound images of human subjects.

Ultrasound has been proposed for aiding epidural needle insertion, but challenges remain in differentiating spinal structures due to noise, artifacts, and inexperience by anesthesiologists in ultrasound interpretation. Moreover, the anesthesiologist needs to measure relevant distances while preserving sterile conditions; therefore, interaction with the ultrasound controls must be minimal. Automated measurement is needed. Beam-steered ultrasound images are captured and spatial compounding is used to improve image quality. Phase symmetry is used to enhance bone (lamina) and ligamentum flavum (LF) ridges. A lamina template is matched to this ridge map using Pearson's cross-correlation, and the most likely lamina positions are found. Then, the lamina is traversed using a LF template with the Pearson's cross-correlation, and the location of the LF is obtained. Tests are performed on 39 sets of compounded ultrasound images in the L2-3 and L3-4 levels of the spine in the paramedian plane. The proposed algorithm can detect the laminas in 38 of the 39 images, and the LF in 34 of the 39 images. In successful detections, the automatic detections versus manual segmentation has an rms error of 0.64 mm and average error 0.04 mm, versus independent sonographer-measured depth has a root-mean-squared error of 3.7 mm and average error 2.5 mm, and versus the actual needle insertion depth has a root-mean-squared of 5.1 mm and average error -2.8 mm. The computational time is 4.3 s on a typical personal computer. The accuracy, reliability, and speed suggest this method may be valuable for helping guide epidurals in conjunction with the traditional loss-of-resistance method.