CenterForDomainSpecificComputing-Poster
Download as PPT, PDF0 likes390 views
This document summarizes work on implementing and accelerating a 3D front propagation segmentation algorithm to measure tumor volumes from medical images. Key points: - The algorithm uses async/finish in Habanero C to spawn parallel tasks that trace contours on individual 2D slices for 3D image segmentation. - Initial results show speedups of 1.38x on a dual core for a 128x128x11 image and 1.26x for a 512x512x29 image. - Future work includes rewriting sequential distance calculation and contour tracing for more speedup, improving seed point detection, and excluding non-nodule regions from segmentation.
CenterForDomainSpecificComputing-Poster
- 1. Acceleration of the Front Propagation Segmentation Algorithm: Towards Tumor Volumetrics Habanero Multicore Software Project Yunming Zhang, Vincent Cave, Vivek Sarkar, Zoran Budimlic Alex A.T. Bui Introduction Rice University Dept of Computer Science Habanero C Results Results For years, the accepted approach to assessing a cancer patient’s response to treatment based on a measurement of the greatest diameter called the Response Evaluation Criteria for Solid Tumors (RECIST). This gross measure neither takes advantage of the information provided by increasingly sophisticated imaging techniques nor incorporates information across multiple image slices. We present initial work on implementing and accelerating a three-dimensional front propagation segmentation algorithm described by J.A. Sethian in “A fast marching level set method for monotonically advancing fronts.” This poster describes the implementation process, initial benchmark results, and challenges encountered. • The 3D segmentation algorithm uses async and finish to spawn tasks operating on different data in parallel. Each task traces the contour on an individual 2D slice of the 3D image. • In Habanero C, async<statement>, causes the parent task to fork a new child task that executes the task defined inside statement. The task returns immediately once finished. In order to synchronize the parallel tasks, the finish <statement> is used to perform a join operation that causes the parent task to execute statement and then wait until all the tasks created within statement have terminate. Algorithm • The habanero-C run time uses a work stealing scheduler to dynamically map tasks to processors. The run time supports work-first and help-first policies along with places for locality. Distance Calculation 1.The algorithm starts by extending the image by one pixel in every direction. 2.From a user-defined seed point, the algorithm expands in all directions in three dimensions based on Dijkstra’s algorithm, where every pixel is regarded as a node in a graph and adjacent pixels are neighbors in the graph. Every iteration, a point closest to the current contour is selected based on its grey value difference to points on the edge of the current contour. 3.The expansions stops at any point that has a grey value change greater than a given threshold, which is decided empirically and interactively. Discussions d Figure 4. Some results from the segmentation and contour tracing algorithm on actual medical image of patient’s lungs scan. Parallelization Parallelization Figure 1. Extending the original image Figure 2. Expansion from the seed point(1,1,0). Numbers are the order. Contour Tracing 1.Given a binary image, traces contours that outline all regions in the 2D image point by point. 2.If multiple regions are present, the algorithm will try to select a region based on its proximity to the seed point and its area. A naïve parallel implementation of the algorithm is written using Habanero C. Contour tracing was identified as a computationally-expensive step in the algorithm. As a result, work was done to redesign the algorithm so that no longer has shared memory access between the tracing of each slice (i.e., the original algorithm was heavily dependent on global variables). In the modified algorithm, tracing of different slices is now executed in parallel to achieve maximum speed up. Dimensions(pixels) Single core(sec) Dual Core(sec) Speed Up 128 * 128 * 11 Figure 3. Contours traced from the given images. 56 0.11 0.08 1.38 512 * 512 * 8 1.25 0.99 4.45 3.52 1.26 Discussions Some future improvements include: •Rewriting the distance calculation and contour tracing algorithms, which are inherently sequential, for further speed up. •Determining an approach for estimating the tolerance value •Automating the detection of the seed point and improvement of the algorithm to exclude regions of the image that are not part of the nodule (e.g., exclude the chest wall). •Achieving additional speed up by targeting additional hardware platforms. 1.26 512 * 512 * 29 In this project, we ported the original java implementation of the algorithm into Habanero C, which required rewriting several library classes. We introduced an improved heap implementation to achieve better performance. we also improved the contour tracing algorithm. On significant challenge is adapting this general-purpose algorithm for use in the medical domain. We tested the algorithm on a variety of chest computed tomography (CT) images with abnormal nodules. While the algorithm was able to accurately capture areas where the boundary of the nodule is distinct (e.g., areas where the nodule is surrounded by air), the segmentation would often bleed into surrounding areas when the nodule was attached to the chest wall. This work is funded by the National Science Foundation Expeditions in Computing Award (NSF CCF-0926127).