PURPOSE

MR provides better tissue contrast for delineating tumor margins and is used for targeted radiation dose-delivery in MR-LINAC systems. Although MR-LINAC systems^1-4 can track tumor targets in real-time and can adapt to targeted dose delivery based on patient breathing, the technology is expensive for cancer treatment, and requires an entirely different system. A cost-effective alternative to MR-LINAC systems is the proposed MR-ultrasound image guided radiation therapy (IgRT) solution that is able to track tumor targets using simultaneous 4D ultrasound (U/S) and MRI^5,6. The MR images acquired prior to radiation treatment under U/S guidance provides high contrast tumor visualization during active radiation treatment guidance on the LINAC.

METHOD

An MR-compatible 4D ultrasound probe was developed to allow hands-free, simultaneous MR-ultrasound image acquisition^5,6. Simultaneously acquired pre-treatment MR and hands-free U/S associates a pair of MR image and U/S volume to each respiratory state over multiple respiratory cycles. The U/S images are used to determine the respiratory states, and the associated MR images that represent that respiratory state in real-time, without requiring physical MR image acquisition during the radiation treatment LINAC phase. To determine respiratory states, displacements of an endogenous fiducial marker, such as a blood vessel or vessel bifurcation in liver were tracked over a time sequence of U/S volumes using a fast and efficient block-matching method⁷. The displacements spanned the space of motion of the endogenous marker due to patient breathing and were used to automatically label the different respiratory states using a hierarchical clustering method^8,9. Vector cosine distance of displacement vectors was used to derive cluster (respiratory state) labels as beam-space data in U/S do not conform to cartesian space. The cluster that occurs most often was hypothesized to signify the end of inspiration or expiration during the respiratory cycle. Because internal organs are moving least at these times, this cluster represents a good opportunity for LINAC dose delivery. All studies were conducted in a GE SIGNA MR750 or a Premier 3.0T MRI scanners with the MR-compatible ultrasound probe driven by a GE Vivid E95 ultrasound scanner. Simultaneous ultrasound and MR data were streamed to an Intel Xeon workstation (512 GB RAM and NVIDIA GeForce GTX Titan X GPU). TTL signals from the MR scanner indicating the start and end of data acquisition for each slice location in conjunction with U/S time-stamps allowed matching of MR images to 3D ultrasound volumes at each time point (Fig. 1). Both 2D MR fast gradient echo (FGRE) and 4D ultrasound images were acquired at 4 fps. Four healthy volunteers were consented under IRB-approved protocols. The U/S probe was placed on the right lateral abdominal wall and secured in place with a strap to image the liver. Two imaging sessions were conducted for the test-retest analysis in order to simulate the patient and probe re-positioning scenarios between pre-therapy and therapy procedures, i.e., the IgRT patient workflow. The fiducial tracking and clustering methods were applied independently on the two U/S volume datasets from the two sessions. Matched MR images corresponding to the end-expiration cluster labels and from the two imaging sessions were aligned using a rigid registration to evaluate the sensitivity of the tracking and respiratory clustering methods subject to patient and probe re-positioning. A total of 5 landmarks on the liver contour were used to drive the rigid registration, while 3 other fiducials within the liver were used to evaluate the accuracy of localizing the target in the expiration MR images of two independent imaging sessions.

RESULTS

Reproducible end-expiration states were obtained automatically using the tracking and clustering methods in both imaging sessions. Fig. 2 shows the end-expiration states with cluster label ‘0’ in both the test-retest data plotted along a one-minute acquisition of U/S (~240 volumes). Fig. 3(a) and 3(b) show the difference images between the matched MR images of expiration states for test and retest data respectively for one subject. Fig. 4(a) and 4(b) shows the difference MR image between the expiration states of test and retest data before alignment and after rigid alignment respectively. Fig. 5(a) shows the fiducials used for computing the rigid transform and 5(b) shows the targets used to evaluate the target localization accuracy between expiration states. A mean value of 1.6 mm ± 1.4 mm for the Euclidean distances of 3 fiducials within the liver was achieved between the test-retest expiration states.

DISCUSSIONS AND CONCLUSIONS

The clustering method consistently captured the end-expiration states in test-retest data that involved repositioning the volunteer and the probe. The error of target localization has been limited to <2 mm as required for reducing the planning treatment volume in IgRT. Further experiments will involve multiple subjects, and finally the cluster labels will be used to train a machine learning model to predict the end-respiratory state labels in real-time during treatment. The results indicate that the MR/ultrasound solution can provide the necessary image guidance for radiation treatment and can be used on any existing MR and LINAC systems, obviating the need for expensive new MR-LINAC system.

Acknowledgements

Funding support: NIH R01CA190298.