In the initial deployment of the hospital compute cluster, only four scanners were con- nected, and the cluster was only used for T2 Shuffling knee exams. This section investi- gates the possible tradeoffs involved with serving many exams from many scanners using compute clusters, if the current distributed system and reconstruction method were used for all exams from these scanners. This serves as an upper bound on the amount of com- putation over the course of the day.
Cleanup 1.13 0.07 1.34
Total 91.85 7.02 108.68
Figure 4.6 shows the MRI scanner schedule from one day out of the one-month study at Lucile Packard Children’s Hospital. Each row represents an MRI scanner, and each entry shows the duration of a scan. The dotted box shows a zoomed-in portion from the schedule to better visualize the individual scans. With the exception of one scanner used in the Emergency Department and reserved for emergencies, there is high utilization across the scanners.
Figure 4.6: MRI scanner schedule from one day out of the one-month study. Each row represents an MRI scanner, and each entry shows the duration of a scan. The dotted box shows a zoomed-in portion from the schedule to better visualize the individual scans.
Using this data, a simulation is performed for the case where one or more compute clusters are reconstructing images from the scanners. To do so, the image sizes are used
to estimate the processing times if they were all to employ PICS-based acquisitions and reconstructions. The cost of the distributed primitives in the cluster were estimated by scaling the measured values from the experiments by the total size of the volume com- pared to the experimental data. For simplicity, it was assumed that the pre-processing and post-processing times scaled as O (NxNyNzlog(NxNyNz)), and that the PICS time scaled
as O (NxNyNzlog(NyNz)); that is, the processing is dominated by the FFT factors and lin-
early scales with the number of concurrent slices. Constants for each stage were fit to match the processing time for the T2 Shuffling knee dataset (Table 4.2).
The simpy discrete event simulation library was used to model contention on the clus- ter. The model has one process representing the hospital, which creates a new reconstruc- tion process for each scan reconstruction task as they arrive. Each scan reconstruction task tries to access a priority resource representing the clusters. Earlier scans are given higher priority. Once a cluster resource is acquired, the reconstruction process holds the resource for a number of time units equal to our run-time estimates for that particular scan. The time spent waiting and reconstructing are logged for each task throughout the simulated day.
1
2
4
8
1
2
4
1
2
1
Nodes/cluster
10
−110
010
110
210
310
4Tim
e (
s)
Single cluster
2 clusters
4 clusters
8 clust.
Avg. Wait Time
Avg. Total Time
Ma . Total Time
Figure 4.7: Simulated average total reconstruction time and average wait time running recon- struction tasks for all scanners in a large hospital over the course of one day, for different cluster configurations
Figure 4.7 shows the result of the simulation. The average wait time is the average time that each reconstruction task spends waiting for the cluster resource. Average total time is the average compute time plus the wait time. The average wait time is reduced
4.6
Conclusion
Developments to PICS-based acquisitions have shown great promise at reducing MRI scan times, but the reconstruction times have remained a barrier to clinical adoption. This chapter explored a distributed reconstruction architecture for T2 Shuffling that achieved clinically feasible latency times. The system was deployed at Lucile Packard Children’s Hospital and used for pediatric knee imaging. The distributed reconstruction approach is well-suited for other PICS acquisitions in which the slices can be independently recon- structed, and it presents an avenue for other reconstruction formulations to benefit from the distributed optimizations.
The optimization of the pipeline led to an overall improvement in performance by 5.3× after speedups and adding additional operations to improve image quality. A 3× speedup came from single-node improvements and a 2.1× speedup came from dis- tributed parallelization on a 4-node cluster. By enabling clinically feasible reconstruction times, a larger number of clinical protocols could potentially leverage PICS acquisitions, leading to shorter overall exam times. Thus, scanner utilization should improve as more patients can be accommodated into the same schedule. Dedicated compute clusters could support this increased utilization without the need for transferring data outside the hos- pital network.
Chapter 5
Clinical Validation
5.1
Introduction
MRI is the predominant imaging method for evaluating musculoskeletal (MSK) pathol- ogy based on its ability to provide superior soft tissue contrast and lack of ionizing ra- diation. In children and adolescents, the knee is the most common joint assessed with MRI [123]. In very young pediatric patients, knee MRI is often used to evaluate devel- opmental conditions and non-traumatic pain [124]. However, the often lengthy exam times continue to be challenging for the pediatric population, compounded by other fac- tors such as patient anxiety and discomfort. Efforts to optimize pediatric MRI protocols aim to maximize information acquired and minimize scan time, often working against an unpredictable deadline when cooperation of the pediatric patient is exceeded [125, 126]. Sedation of the pediatric patient increases the likelihood of a high quality scan, but carries associated risks, including increasing concern for anesthesia associated neurocog- nitive impairment [127]. Moreover, the need for an MRI exam under sedation requires a complex multidisciplinary approach [128] and can decrease the overall efficiency of a ra- diology department [129]. A secondary effect of the long exams is high cost, limiting the value of the MRI exam in the context of clinical care. Nevertheless, the basic conventional 2D knee MRI protocol still requires multiple acquisitions in axial, coronal, and sagittal planes due to voxel anisotropy, and also includes sequences with different contrasts to optimize assessment of various structures and pathology [125, 130]. A single-sequence MRI protocol can potentially alleviate some of the difficulties encountered in pediatric MRI.
Volumetric variants of FSE [3, 40] have been described to simplify and accelerate joint MRI based on potential reformattability into multiple planes [131–133]. However in clini- cal practice, standard 2D imaging still predominates largely due to blurring of the vol- umetric images. This is likely due to long echo trains required for scan efficiency of volumetric imaging, which leads to image blurring due to T2 decay [4, 40]. This work evaluates T2 Shuffling [117] as an alternative volumetric FSE acquisition due to the re-
suffice as a single-sequence rapid pediatric knee protocol. In Section 5.2, T2 Shuffling is compared to the routine knee MRI protocol for image quality and diagnostic perfor- mance. Building on the evidence that missing clinically relevant pathology is unlikely, Section 5.3 evaluates a targeted knee MRI exam based on T2 Shuffling with significantly shorter exam times. As the table time is significantly reduced compared to the conven- tional protocol, patients are scheduled for the targeted exam between existing slots. The targeted knee MRI is billed with a reduced charge modifier code, lowering the cost by a factor of three compared to the standard lower extremity joint MRI. As additional patients can now be accommodated into the MRI schedule, the cost of the MRI is simply profes- sional costs of the radiologist and minimal costs associated with reporting infrastructure. The magnet, technologists, and receptionists are already present. A significant additional benefit is same-day scheduling, which reduces lost productivity of parents and children at their work and school.