SU-G-BRA-02: Development of a Learning Based Block Matching Algorithm for Ultrasound Tracking in Radiotherapy




To develop an ultrasound learning-based tracking algorithm with the potential to provide real-time motion traces of anatomy-based fiducials that may aid in the effective delivery of external beam radiation.


The algorithm was developed in Matlab R2015a and consists of two main stages: reference frame selection, and localized block matching. Immediately following frame acquisition, a normalized cross-correlation (NCC) similarity metric is used to determine a reference frame most similar to the current frame from a series of training set images that were acquired during a pretreatment scan. Segmented features in the reference frame provide the basis for the localized block matching to determine the feature locations in the current frame. The boundary points of the reference frame segmentation are used as the initial locations for the block matching and NCC is used to find the most similar block in the current frame. The best matched block locations in the current frame comprise the updated feature boundary. The algorithm was tested using five features from two sets of ultrasound patient data obtained from MICCAI 2014 CLUST. Due to the lack of a training set associated with the image sequences, the first 200 frames of the image sets were considered a valid training set for preliminary testing, and tracking was performed over the remaining frames.


Tracking of the five vessel features resulted in an average tracking error of 1.21 mm relative to predefined annotations. The average analysis rate was 15.7 FPS with analysis for one of the two patients reaching real-time speeds. Computations were performed on an i5-3230M at 2.60 GHz.


Preliminary tests show tracking errors comparable with similar algorithms at close to real-time speeds. Extension of the work onto a GPU platform has the potential to achieve real-time performance, making tracking for therapy applications a feasible option.

This work is partially funded by NIH grant R01CA190298