Synopsis

Keywords: Machine Learning/Artificial Intelligence, MSK

In clinical practice, oblique coronal and sagittal scans have been shown to be superior for visualization of ACL and assessment of ACL tears. In this work, we demonstrate a Deep Learning based method to obtain these complex scan planes (Coronal ACL and Coronal Blumensaat) directly from standard 2D tri-planar Localizers. Results on 487 test volumes show <0.6mm mean error for both planes, indicating suitability for clinical usage.

Summary of Main Findings

A Deep Learning (DL) model adapted for Coronal Knee MR prescription on standard 2D tri-planar Localizer images for ACL and Blumensaat Line was obtained with mean error < 0.6mm. Results indicate suitability for clinical usage of ACL prescription.

Introduction

Oblique coronal ACL imaging in addition to oblique sagittal knee plane is used for improving the diagnostic accuracy of detecting single/double bundle injury of ACL¹ and improving detection of partial/complete ACL tears^2,3. As ACL has fine anatomy (thickness < 5mm), precise angulation is challenging and requires highly skilled technicians in clinical practice⁴. In this work, we present an Intelligent Slice Placement (ISP) framework using standard 2D tri-planar Localizer images that allows for contiguous visualization of ACL, irrespective of changes in patient position, minor anatomical changes, and technologist training. Previous efforts to automate prescription include use of specialized 3D localizers and landmark detection, or more recently, use of DL-based generalized framework for MRI scan plane prescription^5,6. In this work, we have adapted the previously described generic ISP framework for Brain⁶ and Knee⁷ with only training data and plane definitions adapted for ACL in the oblique coronal plane.
Traditional algorithms use high-resolution data to visualize, segment and generate ACL plane prescription as ACL is not easily visualized in standard low-resolution 2D Localizers. Algorithms that rely on explicit ACL computation from high-resolution data produce inaccurate results in cases where ACL is torn, attenuated, or obscured by artifacts from metal implants. The clinically accepted alternative in such scenarios is to use the bone-centric Blumensaat Line as a surrogate for ACL plane⁸.
In our work, we use a DL network to predict ACL plane and Blumensaat Line from standard Localizer images without explicit segmentation of the structures by predicting the scan plane masks. Even when the integrity of ACL is not preserved or the structure is not visible, we hypothesize that the DL network is able to learn the entire 3-dimensionalanatomical context and produce robust plane predictions. Our DL models simultaneously predict both ACL and Blumensaat planes and hence allow both options for oblique coronal ACL prescription based on clinician preferences.

Methods

Subjects: Knee MRI data for study came from multiple sites. A total of 1306 knee exams from volunteers as well as patients were included in the study. All the studies were approved by respective Institutional Review Boards.
MRI Scanner and Data Acquisition: Localizer Data was acquired on multiple MRI scanners (GE 3T Discovery MR 750w, GE Signa HDxt 1.5T, GE 1.5T Optima 450w) and with different coil configurations (e.g.,16-/ 18-channel TR Knee coil, GEM Flex coil etc.). Localizer data varied in contrast (FGRE, SSFSE), image resolution, field of view, slice coverage and matrix size across subjects.
DL Methodology: Ground-truth (GT) landmark points were marked by a trained radiologist and then translated into imaging planes for DL segmentation. Two points each were marked on the ACL and Blumensaat Line to generate the tilt of oblique coronal planes (Fig.1). A total of 1062 exams (30,652 volumes after augmentation) were used for training/validation and 258 exams (487 volumes after augmentation) were used for testing. A variety of Intensity and Geometry augmentations were applied. Multi-task Learning was used to train single segmentation model that infers on sagittal Localizer images to output two segmentation plane masks – one each for ACL and Blumensaat Line (Fig.2).
Accuracy Assessment: Accuracy was assessed by calculating mean absolute distance (MAD) error and angle error between GT and DL predictions for both planes. MAD error < 1 mm and angle-error < 4⁰ was considered as acceptable for ISP.

Results and Discussion

Mean MAD error obtained was 0.55mm for ACL plane and 0.48mm for Blumensaat Line plane predictions on test cohort of 487 datasets on comparison with GT (Fig.3). Apart from clinically normal cases, the test cohort also contained Localizers that include pediatric, geriatric cases, fractured knees, osteotomies, pathologies, and metal implants. MAD errors for selective subset of the test cohort that include the above are as shown in Fig.3.
The model provided robust plane predictions, even in presence of metal (Fig.4) and torn/obscured ACL (Fig.5). Effectiveness of our ISP framework is demonstrated in a prospective reformatting of high-resolution isotropic CUBE T2-weighted knee using predicted planes (Fig.5), which indicates contiguous visualization of ACL.

Conclusion

We have adapted a generalized DL-based intelligent slice placement framework for predicting the oblique coronal planes along ACL and Blumensaat Line. The results indicate that the framework allows for successful oblique coronal ACL prescription with <0.6mm error and <4 degrees error. Results demonstrate its clinical readiness with robust performance even in presence of artifacts in the Localizer data. Bone-based Blumensaat Line landmark can be successfully used as surrogate for ACL, in cases with severe ACL injury (torn, attenuated) or obscured ACL.

Acknowledgements

No acknowledgement found.