Synopsis

Keywords: Machine Learning/Artificial Intelligence, Visualization, Scan Planning, Spine, Reformatting

We present a generalized DL-based intelligent slice placement framework for planes of cervical neural foramina (CF) and lumbar pars interarticularis (PI) for spine MRI. CF and PI scan improves assessment of foraminal stenosis and lumbar spondylolysis respectively, but requires highly skilled operator for accurate prescription. Our approach enables automatic patient-specific scan plane prescription from routine axial T2W images. In our test, it achieves mean error of <0.7 mm and <0.2 degrees and demonstrates similar/better contiguous anatomical visualization to manual scans on retrospective reformatting of 3D data. These results indicate that our approach is accurate and suitable for clinical usage.

Introduction

Deep learning (DL) based intelligent scan plane (ISP) prescription has demonstrated good repeatability and improved performance vis-à-vis manual prescription for various landmarks across anatomies^1-4. In this work, we have leveraged a similar approach for automating the prescription of oblique planes across the long axis of cervical neural foramina (CF) and oblique plane along lumbar pars interarticularis (PI). The oblique CF scan plane perpendicular to the long axis of foramen provides best assessment of cross-sectional area reducing inter-observer variabilities for assessment of foraminal stenosis compared to the standard sagittal or axial images of the spine^5,6. Similarly, oblique PI plane improves assessment of pars interarticularis defects in lumbar spine as compared to the sagittal and axial planes alone⁷. Automated prescription of these scan-planes would be very impactful to reduce variability in acquisition and setup time, irrespective of technologist familiarity with spine anatomy. Compared to previous DL-based ISP methods^1-3, we have utilized regular T2w axial images, instead of localizers.

Methods

Data acquisition: 454 cervical and lumbar spine exams (IRB approved) from two clinical sites were used, which included axial 2D T2w spine images (multi-slice multi-angle (MSMA) or block acquisitions) over multiple vertebrae. Data came from various GE 1.5T and 3.0T MRI scanners with varying protocol parameters. Additionally in two volunteers, a trained technologist acquired 3D sagittal left and right cervical foramina data (3D CUBE, TR:2000, TE: 90.7, 210x210x40mm³ FOV, 0.2x0.20.5mm³ resolution) as well as a 3D axial data stack (3D CUBE, TE: 90.3, TR: 3479, 180x180x80mm³ FOV, 0.35x0.35x1mm³ resolution). The plane prediction was done on 3D axial stack and was reformatted accordingly using the predicted plane.
Ground-truth (GT) marking: A trained radiologist marked the plane of CF and PI on both the left and right side (with different labels) of the landmark planes (Fig.1).
DL Data Preparation: To account for variations in slice thickness and angulations from MSMA stack, the GT marking was extended to neighboring slices using 1D dilation along slice direction (Fig.1B). The data was then cropped to retain only the center region of the axial image. All data was resampled to 256x256 matrix size and several intensity-augmentation were applied, including coil sensitivity, smoothening and sharpening filters etc. SimpleITK⁸ was used for data preparation and post-processing.
DL Methodology: For CF, a total of 116 Train (1326 augmented volumes), 18 Validation (208 augmented volumes) and 19 test subjects were available. For PI, a total of 200 Train (818 augmented volumes), 28 validation (117 augmented volumes) and 85 test cases were available. For segmentation of CF and PI scan planes, a variant of U-Net architecture¹, was adapted with four layers of dyadic reduction and expansion with skip connections. The loss function was a combination of dice coefficient and boundary distance loss⁹. The dice loss function was the primary loss function till the boundary loss was >0 and then combined with boundary loss using weight factors as: 0.33 dice, 0.67 : Boundary loss.
Accuracy Assessment: The trained model predicted binary mask for CF and PI scan planes which comprises both the left and right planes. These were separated using image center information to obtain separate left and right scan planes. Analytical form of the scan planes were obtained by fitting planes to the predicted scan plane mask. Accuracy was assessed by calculating mean absolute distance (MAD) error and angle error between GT and DL-predicted planes for all the landmarks. MAD error <1 mm and angle-error <3⁰ was considered as acceptable for ISP.
We also retrospectively reformatted the 3D axial CUBE data using the predicted CF scan plane for left and right directions to generate sagittal view and compared it to manually prescribed sagittal CF plane prescriptions. The results were reviewed with a radiologist to ascertain the clinical acceptance.

Results and Discussion

For DL based CF prediction model, the angle and MAD errors for both CF and PI planes along right and left direction were within the acceptable limits (Fig 2). Fig 3 shows the prediction of CF plane on the two volunteers datasets. The DL-predicted planes were used to fit analytical plane for reformatting the isotropic axial 3D data. Fig 4 and 5 demonstrate the results of reformat and compare to manually prescribed CF plane. We notice that in both the cases, the DL-based CF plane reformatted data is similar or slightly better in providing a contiguous visualization of foramina; especially with a straight spine. In case of curved spine (Fig 5), a single plane will not be able to fit the entire foramina region and this is evident in both the manual and DL-based plane prescription. However, in clinical practice, the foramina plane will be typically acquired around C4-C5 region (and similarly L4-L5 region for PI plane) and hence results are acceptable for clinical prescription.

Conclusion

We adapted a generalized DL-based intelligent slice placement framework for automated planning of cervical foramina and pars interarticularis for MRI spine exams. It achieved mean error of <0.7 mm and <0.2 degrees in our test. Prospective reformatting of 3D data demonstrates similar/better contiguous visualization of anatomical structure as compared to manual scan. These results indicate that our framework allows for patient specific automated plane prescription, which can be used in clinical practice.

Acknowledgements

No acknowledgement found.