Purpose: This study aims to quickly and accurately reconstruct multiple needles in 3D ultrasound (US) images using an unsupervised dictionary learning on the images without needles for ultrasound-guided prostate brachytherapy.
Methods: Our approach employs the available 3D US images without needles to learn a sparse dictionary with the proposed order-graph regularized dictionary learning algorithm. The resultant dictionary transforms the extracted background structure of US image to reconstruct the US image corresponding to target US images with needles. We achieve the residual images of which the pixels implicitly corresponds to needles. Then our needle detection algorithm clusters the residual pixels in term of their intensities to product regions of interest on the cluster centers, followed by repeatedly performing random sample consensus algorithm with fitting a line. Rely on the fitted lines, the detection results are further refined by removing unreasonable needle points and correcting their locations at all slices. Finally, we draw out the points as needle tips, where the intensities along the needle drop sharply and the intensities of their nearby voxels on both sides are uniform.
Results: Experiments are conducted on 3D US images from 40 patients without needles and 8 patients with needles. The visualization results of needle points over each slice show that the proposed method can label the needles at correct location except in a few cases. The results of tip detection show that most of tips can be precisely captured while the average error can be limited into 1mm.
Conclusion: The proposed dictionary-learning method can effectively utilize the background structure and the spatial continuity of a needle to correctly detect multiple needles from noisy US images simultaneously while also delivering acceptable error for tip detection. This technique could provide accurate needle reconstruction for US-guided HDR prostate brachytherapy, greatly facilitating the routine clinical workflow.
Not Applicable / None Entered.