A parallel, pipeline-based online analysis system for Interaction Vertex Imaging

TL;DR

The paper addresses the need for clinically relevant online range monitoring in carbon ion radiotherapy by developing a parallel, pipeline-based analysis system for Interaction Vertex Imaging (IVI). It implements a multi-stage, task-parallel workflow—from trigger event formation through vertex reconstruction and range monitoring—coupled with parallel clustering to achieve sub-200 ms end-to-end latency for range shift detection. Profiling shows clustering as the dominant compute cost, with performance improved by distributing work across multiple clustering workers and optimizing inter-thread communication. The results demonstrate that online IVI can provide timely feedback during beam delivery, enabling treatment pauses or aborts when range errors are detected, and thus potentially reducing dose to healthy tissue while maintaining tumor control.

Abstract

Objective Interaction vertex imaging (IVI) is used for range monitoring in carbon ion radiotherapy, detecting depth differences between Bragg peak positions. Online range monitoring, which provides feedback during beam delivery, is particularly desirable, creating an opportunity to detect range errors before the treatment fraction is completed. Incorporating online range monitoring into clinical workflows may therefore improve the safety and consistency of radiotherapy. Approach The data analysis system was broken into a task-parallel pipeline approach, to allow multiple analysis stages to occur concurrently, beginning during acquisition. Computationally-expensive operations were further parallelized to reduce bottleneck effects. Data collected from irradiation of homogeneous plastic phantoms was replayed at the same rate it was initially acquired, to mimic data acquisition, and the time required to determine a range shift was measured. Main Results With an optimized pipeline, the delay between the end of irradiation and the determination of a range shift is consistently less than 200 ms. The majority of this time is associated with the final range shift determination, with a minor effect from the time required to analyze the last data packet. The most significant contribution to an optimized analysis workflow is the formation of clusters, requiring almost 50% of compute time. Significance This system is the first IVI implementation to achieve clinically-relevant online analysis times. The 200 ms time required to determine a range shift is less than the time required to accelerate a new spill in a synchrotron, and is comparable to the time required for reacceleration if multiple energies are delivered in the same spill. Clinical implementation of online range monitoring would allow treatment to be quickly paused or aborted if significant range errors are detected.

Figures (8)

• Figure 1: Schematic of data flow within the data acquisition system. Each GBTxEMU readout board (green) aggregates data from two front-end boards (blue), which are each responsible for reading out an entire sensor side. Aggregated data from the GBTxEMU is sent along an optical link to the GERI readout interface (magenta), which combines the entire system into a single datastream.
• Figure 2: Schematic of the data analysis pipeline. Each uplink is processed in its own thread (blue). All links corresponding to the same side of a sensor are combined at the clustering phase (grey), where groups of adjacent and coincident trigger events are formed. Coincident clusters from opposite sides of the same sensor form hits (red), and coincident hits from different layers of the tracker are used to form tracks (cyan). Each track is then independently evaluated using the filtered IVI method (magenta); the resultant interaction vertices are used in the range shift determination process (yellow). Each element represents a single thread of execution; the counts are reflective of the evaluated configuration. However, the system is highly flexible, and each element (with the exception of hit formation) is able to accept an arbitrary number of inputs.
• Figure 3: Representative example of the sequence of data in a single elink. TS_MSB frames (green) are automatically generated at regular intervals to record the passage of time. When Hit frames (yellow) are ready to be sent, Hit-associated TS_MSB frames (blue) are also generated if necessary. These Hit-associated frames are only required if the most recent TS_MSB frame sent over the link does not match the most significant bits of the upcoming Hit. Three cases are represented in this figure. In the upper right, the Hit-associated TS_MSB frame is clearly not a part of the automatically-generated sequence. In the lower left, the Hit-associated TS_MSB frame could be a part of the automatically-generated sequence, but it is clearly distinguished by being followed by Hit frames. Once the Hit frames have been sent, the automatically-generated timestamp is reasserted over the link. In the lower right, no Hit-associated TS_MSB frame is generated, as the automatically-generated TS_MSB frame is already valid for these Hits. In this case, the automatically-generated TS_MSB frame is treated as a Hit-associated frame, so no additional TS_MSB frame is required, and the passage of time resumes with the next automatically-generated frame.
• Figure 4: Schematic of data flow in the parallel clustering algorithm, which is a replacement for the serial grey element from Figure \ref{['fig:pipeline']}. Data from uplinks (blue) are first processed by a single thread (green), before being distributed to a pool of worker threads (grey), with each thread dedicated to a separate portion of the sensor area. Clusters formed by the worker threads are then merged by a single thread (orange) to form a single output data stream for the next stage of the pipeline. The remaining analysis then continues as shown in Figure \ref{['fig:pipeline']}.
• Figure 5: Analysis time as a function of worker count per sensor side in the clustering phase. The data points at zero workers represent the serial model. The presented data is for a sleep time of 30. A line has been drawn between points for each data series, to guide the eye. The parallel model yields reduced performance relative to the serial for low worker counts. At very high worker counts, contention effects reduce the advantage of parallelism, particularly at high data rates.