3832

Interleaved White Marker Contrast with bSSFP Real-Time Imaging for Deep Learning based Needle Localization in MR-Guided Percutaneous Interventions

Jonathan Weine^1,2, Rainer Schneider¹, Urte Kägebein³, Bennet Hensen³, Frank Wacker³, and Florian Maier¹

¹Siemens Healthcare, Erlangen, Germany, ²TU Dortmund, Dortmund, Germany, ³Hannover Medical School, Hannover, Germany

Synopsis

Automatic localization of needles in real-time images can facilitate MR-guided percutaneous interventions. It enables automatic slice repositioning and targeting support and, thus, allows for faster workflows. The improvement of deep learning based passive needle tracking by using both, anatomical and positive contrast images as input was investigated. A prototype bSSFP sequence for interleaved acquisition of k-space lines for conventional and positive contrast with Cartesian readout was implemented and evaluated ex-vivo and in-vivo. The U-Net segmentation algorithm showed superior performance when using both contrasts. In conclusion, this method is a promising approach for robust needle localization in real-time interventional workflows.

Introduction

Automatic needle localization can facilitate MR-guided interventions by enabling automatic slice repositioning and targeting support¹ . Several approaches such as active tracking², optical tracking^3-5 or classical image processing⁶ on additional contrasts^7,8 have been proposed. Recently, the utilization of deep learning (DL) algorithms for needle artifact segmentation showed promising results for detecting needle artifacts^9,10. For segmenting single 2D images, structures like banding artifacts and physiological signal voids can cause ambiguities in detection. In this work, the feasibility of combining white marker contrast and DL-based needle artifact segmentation to improve needle localization during MR-guided interventions is investigated.

Methods

A prototype real-time bSSFP sequence was extended, to enable an interleaved acquisition of k-space lines for conventional and white marker contrast. Dephasing was implemented in slice selection and read-out direction to acquire the white marker signal for in-plane orientation of needles as well as orthogonal orientation (cf. Fig. 1). Flip angle was set to 70°, TE=2.06ms (TR=2TE) and bandwidth to 555 Hz/pixel. TE had to be extended depending on the applied dephasing moment due to gradient limitations (up to 2.64ms). A dataset of 700 in-plane and 400 orthogonal images (FOV= 300x300mm², 192x192-matrix) including both contrasts, was acquired on a scanner (MAGNETOM Aera, Siemens Healthcare, Erlangen, Germany) using an ex-vivo phantom (cf. Fig. 2). Seven needle insertions with different trajectories were performed and manually annotated. During the ex-vivo experiments a 19G needle (Cook Medical, Bloomington, IN, USA) was used. Based on the ex-vivo data a non-product U-Net segmentation¹¹ algorithm was trained for detecting the artifact in orthogonal and in-plane images respectively. To investigate the benefits of using the additional contrast as input, the algorithm was trained and tested with and without the second input channel using the same hyperparameters in a 7-fold cross validation scheme. Random augmentation, including rotation, flipping, zooming, shifting and adding Gaussian noise was applied during training to prevent the model from overfitting. The performance was evaluated by calculating F1 scores, number of detected connected structures for orthogonal slices, and angular deviation between the segmented artifact orientations and annotations. Feasibility of acquiring signal from the off-resonant spins close to the needle was evaluated in an in-vivo animal experiment using a 1.5T scanner (MAGNETOM Aera, Siemens Healthcare, Erlangen, Germany) and a 14G microwave ablation needle (MedWaves, Inc., San Diego, United States). The required dephasing was determined by applying a parameter sweep (-2.809 rad/mm to 0 rad/mm)

Results

Fig. 3 shows images of the ex-vivo phantom with both contrasts and the comparison of inference for the images of the trained algorithms.

Orthogonal slice orientation: Signal conservation due to compensation of the dipole field around the needle at some points was clearly visible in the white marker contrast image. These signal conservations form a unique texture that improves separability. Using both contrasts for training reduced the number of test images with multiple structures being classified as needle artifact (Fig. 4a) as well as the number of missing detections (Fig. 4b).

In-plane slice orientation: Segmentations inferred by the algorithm using the additional white marker contrast, show increased F1 scores (Fig 4d), as well as contours that better match the artifact (cf. Fig 3 g,h). Therefore, the segmented areas’ main axis matched the annotated needle orientation more closely (Fig 4c). Figure 5 shows two representative in-vivo images with clearly visible positive signal around the needle and the artifact was spatially coherent over both contrasts under free breathing conditions.

Discussion

Utilizing positive contrast images improves the F1 score of artifact segmentation for in-plane and orthogonal needle orientation. This enables a more robust 3D localization of needles using DL algorithms. The implementation of the sequence allows for measuring two contrasts using the balanced steady state at each read-out. The pulse sequence can be used interactively and parallel imaging features are fully available. This allows for activating the white marker acquisition only, if the confidence level of localization drops below a certain threshold. This can minimize additionally required acquisition time for the white marker images. As expected, dephasing differed between the different devices. The amount and variability of training data needs to be increased, to allow for a more detailed assessment of the improvement. Therefore, acquisition of additional in-vivo data will be performed.

Conclusion

Using white marker contrast signal as additional input channel for DL algorithms designed to localize needles in percutaneous interventions allows for more robust segmentation of the needle artifact. The presented work might improve the clinical workflow for MR-guided needle interventions, by utilizing the increased robustness in orthogonal needle orientation for automated slice repositioning without the need for any additional hardware accessories

Acknowledgements

The authors thank Inga Brüsch, PhD and Dr. Regina Rumpel for the veterinarian supervision of the in-vivo experiment.

References

1. Barkhausen J, Kahn T, Krombach G A, Kuhl C K, Lotz J, Maintz D, Ricke J, Schönberg S O, Vogl T J, Wacker F K. White Paper: Interventional MRI: Current Status and Potential for Development Considering Economic Perspectives, Part 2: Liver and Other Applications in Oncology. Fortschr. Röntgenstr 2017; 189(11): 1047-1054. DOI:10.1055/s-0043-112336

2. Zimmermann, H. , Müller, S. , Gutmann, B. , Bardenheuer, H. , Melzer, A. , Umathum, R. , Nitz, W. , Semmler, W. and Bock, M., Targeted‐HASTE imaging with automated device tracking for MR‐guided needle interventions in closed‐bore MR systems. Magn. Reson. Med. 2006; 56: 481-488. DOI:10.1002/mrm.20983

3. Kägebein U, Godenschweger F, Armstrong BSR, Rose G, Wacker FK, Speck O, Hensen B. Percutaneous MR-guided interventions using an optical Moiré Phase tracking system: Initial results. PLoS ONE 2018; 13(10): e0205394 DOI:10.1371/journal.pone.0205394

4. Park Y, Elayaperumal S, Daniel B, Ryu S C, Shin M, Savall J, Black R J, Moslehi B, Cutkosky M R. Real-Time Estimation of 3-D Needle Shape and Deflection for MRI-Guided Interventions. IEEE/ASME Transactions on Mechatronics Dec. 2010; 15(6):906-915. DOI:10.1109/TMECH.2010.2080360

5. Busse H, Garnov N, Thörmer G, Zajonz D, Gründer W, Kahn T and Moche M (2010), Flexible add‐on solution for MR image‐guided interventions in a closed‐bore scanner environment. Magn. Reson. Med., 64: 922-928. DOI:10.1002/mrm.22464

6. Rothgang E, Gilson W D, Wacker F, Hornegger J, Lorenz C H and Weiss C R (2013), Rapid freehand MR‐guided percutaneous needle interventions: An image‐based approach to improve workflow and feasibility. J. Magn. Reson. Imaging, 37: 1202-1212. DOI:10.1002/jmri.23894

7. de Oliveira A, Rauschenberg J, Beyersdorff D, Semmler W and Bock M (2008), Automatic passive tracking of an endorectal prostate biopsy device using phase‐only cross‐correlation. Magn. Reson. Med., 59: 1043-1050. DOI:10.1002/mrm.21430

8. Campbell-Washburn A E, Rogers T, Xue H, Hansen M S, Lederman R J and Faranesh A Z. Dual echo positive contrast bSSFP for real-time visualization of passive devices during magnetic resonance guided cardiovascular catheterization. J. Cardiovascular Magn. Reson. 2014; 88(16). DOI:10.1186/s12968-014-0088-7

9. Weine J, Rothgang E, Wacker F, Weiss C R, Maier F. Passive Needle Tracking with Deep Convolutional Neural Nets for MR-Guided Percutaneous Interventions. Proceedings of 12th Interventional MRI Symposium Oct 2018. 12:53.

10. Mehrtash A, Ghafoorian M, Pernelle G, Ziaei A, Heslinga FG, Tuncali K, Fedorov A, Kikinis R, Tempany CM, Wells WM, Abolmaesumi P, Kapur T. Automatic Needle Segmentation and Localization in MRI with 3D Convolutional Neural Networks: Application to MRI-targeted Prostate Biopsy. IEEE Transactions on Medical Imaging, 2018; DOI:10.1109/TMI.2018.2876796

11. Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. MICCAI 2015. Lecture Notes in Computer Science 2015; 9351. Springer, Cham. DOI: 10.1007/978-3-319-24574-4_28

Figures

Fig. 1: Sequence diagram of the interleaved k-space line acquisition. Purple dashed lines indicate SS-dephasing for in-plane needle orientation. Green dashed lines indicate RO-dephasing for orthogonal needles. Acquisition time at least doubles, since gradient strength limitations can cause timing changes. Implementation is compatible with PAT.

Fig. 2: Pictures of the ex-vivo phantom. To get in-vivo like structures an entire porcine liver was placed inside a pork side without further offal.

Fig. 3: a,e) conventional bSSFP contrast b,f) white marker contrast c,g) inference of algorithm trained without white marker contrast d,h) inference of algorithm trained with white marker contrast. Purple area is equal to the segmentation map generated from annotations thresholded at 0.5. F1 score for shown inference thresholded at 0.95 given for comparison. Voxel spacing is 1.67x1.67x10mm³ for all images. Top-row: orthogonal needle orientation, red arrow marks needle artifact, dephasing in RO-direction was -2.622 rad/mm; bottom-row: in-plane needle orientation, dephasing in SS-direction was -2.06 rad/mm. Red arrows mark the tip position for all subfigures.

Fig. 4: Normalized histograms comparing the performance of segmentation algorithm depending. Blue curves correspond to conventional contrast only; orange curves correspond to additional use of white marker contrast. Top row: orthogonal needle orientation a) Number of detected contours in inference at threshold 0.95 b) F1 scores at threshold 0.95, F1 = 0 indicates missing detection; Bottom row: in-plane needle orientation c) Angular difference between main axis of label mask and segmentation at threshold 0.95 d) F1 scores at threshold 0.95, mean increases (0.58 to 0.61), variance decreases (0.186 to 0.180).

Fig. 5: In-vivo images under free breathing, dephasing of -0.936 rad/mm in slice selection direction, voxel spacing 2.7x2.7x10mm³. a,c) conventional bSSFP contrast b,d) white marker contrast. Red arrows mark the same image coordinates.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)

3832