Semi-Automated Segmentation of 3D Medical Ultrasound Images

Link to full document


Automated, accurate 3D segmentation is critical to achieve the full potential of 3D ultrasonic imaging. The goal of this research is to develop a robust automated system for the segmentation (boundary identification) of targets in 3D medical ultrasound images.  This is part of the 3D portable ultrasound system being developed in the WPI Ultrasonics Laboratory under the direction of Dr. Pedersen.  Applications include volume assessment of free fluid volumes of the abdomen, evaluation of cancer treatment efficacy by monitoring changes in tumor volume, and improved measurements in obstetrics.  This will allow doctors in the field to visualize 3D models of internal structures and fluid volumes as well as calculate statistics such as shape, volume, or track those statistics over time for a patient, giving the clinician additional diagnostic tools.  In addition to automatic segmentation, arbitrary 2D scan planes can be extracted from the 3D volume and viewed, providing views of internal structures otherwise impossible to acquire with traditional 2D B-mode ultrasound scans.  As well as being automated, the system should ideally be fast enough to handle the segmentation in real time.

In this segmentation study, we specifically targeted simulated cysts, cysts in tissue-mimicking phantoms, and boundaries of prostates, obtained clinically. Segmentation was performed directly in 3D using the level set method, but required manual initialization and was hence semi-automated.  Various pre-processing techniques were explored to find the most robust and accurate segmentation results. Speckle-type noise, prevalent in ultrasound imaging, was a major concern and was the focus of most of the pre-processing methods.  It was found that evolving the 3D images under their own curvature (curvature flow imaging) resulted in the most accurate segmentation results.

Segmentation was performed by placing an initial 3D surface “within” the image volume and subjecting it to forces that deformed it over time. Balloon, curvature, advection, and a “speed image” were used to supply the forces influencing the evolving front.  Once the contour stopped moving from iteration to iteration (convergence criteria), the final surface was extracted and compared with ground-truth models.

Ground-truth models were generated for targets with known geometries, and obtained from medical doctors for the case of clinical human (prostate) data. A surface-to-surface distance metric was developed to evaluate the RMS distance between points on the segmentations and the ground-truth models. This was repeated for the different pre-processing methods to determine the ones most useful for obtaining accurate boundary detection.

It was found that the methods employed in this research were able to determine the location of target boundaries within the error margin of two doctors hand-segmenting the same objects.  Currently, processing time is an issue for real-time implementation, although there are several avenues to pursue with further development that will enable this to be done in real-time, namely more efficient algorithms for implementing the active contour segmentation, and using dedicated hardware to parallelize the image pre-processing.