Dawei Gui1, Aanchal Mongia2, Chitresh Bhushan3, Jeremy Heinlein1, Kavitha Manickam1, Muhan Shao3, Uday Patil2, and Dattesh Shanbhag2
1GE Healthcare, Waukesha, WI, United States, 2GE Healthcare, Bengaluru, India, 3GE Healthcare, Niskayuna, NY, United States
Synopsis
Keywords: Other AI/ML, Machine Learning/Artificial Intelligence
Motivation: A fully automatic workflow for scan plane prescription is desirable in clinical settings.
Goal(s): Our goal is to demonstrate a deep learning-based MRI scan workflow for fully automated MR scanning in the prostate.
Approach: This new scan workflow will identify anatomical landmarks and scan planes for prostate planning (coverage, FOV and orientation) from coil sensitivity and 3plane scout images.
Results: The deep learning-based anatomy recognition showed acceptable average location error below 5mm and plane orientation error below 10 degrees.
Impact: As no interaction from the operator is required
to complete a full MR prostate scan, it paves the way for fully automated MR
scans for the prostate anatomy.
INTRODUCTION
MRI
is inherently a multi-planar imaging modality. An automatic
workflow for scan plane prescription of different landmarks and anatomies is
desirable in clinical settings to reduce MRI exam time and improve image consistency,
especially in longitudinal studies. Previously, it was demonstrated that
intelligent slice placement (ISP) for multiple landmarks in the brain and knee
is possible using a deep learning-based framework based on standard 2D
tri-planar localizer images [1-3]. However, the scout images are setup manually
in blinded manner and ISP algorithm must accommodate variability in scout
placements for robust performance. Sometimes the scout images are repeated due
to insufficient anatomical coverage. Previously, it was also demonstrated that
the coil sensitivity maps can be used to automatically find the region of
interest of spine anatomy to correctly position the volume for the 3plane scout
scan, typically the first scan for any patient scan planning [4].
In
this work, we demonstrate a generalized, deep learning (DL)-based framework for
a fully automated MRI scan workflow for the prostate anatomy. It consists of automated
volume prescription for the 3plane scout scan using CalLocNet model based on
the 3D coil sensitivity scan to identify the prostate location from the coil sensitivity map volume.
After the 3plane scout scan is acquired, multiple deep neural networks (ScanNets)
will identify the anatomical structures around the prostate in the scout volume,
which include apex and base of the prostate, rectum, pubis, seminal vesicles,
L5 and S1, left and right hip, prostate, and urethra. All the above information
can then be used to prescribe the imaging volume center, field of view, slice
coverage and volume orientation for acquiring the high-resolution prostate
MR scans automatically.METHODS
Data: Cohort
A. 66 coil sensitivity maps and
3plane scout images of volunteers were collected for the CalLocNet. Cohort
B. 741 volume of patients from different hospitals were collected
and used for training and evaluation of ScanNets models. In all training and
evaluation testing, 80% of data were used for training, 10% for validation and
10% for testing.
Models: The
fully automated workflow is supported by 5 deep learning models (based on
architecture for segmentation [6]) as described below:
1) CalLocNet: Identifies
the prostate location from the coil sensitivity map volume. The
prostate location is then used to determine the volume for 3plane scout scan.
2) ScanNets: Four
models to Identify specific
anatomy locations and volumes around the prostate region. The first model (3D
SagLocNet) identifies apex and base of prostate, L5, S1, pubis, rectum, seminal
vesicle. The second (3D CorLocnet) identifies left and right hip, the third one
(2D ProstateSeg) identifies prostate mask and the fourth model (3D Oblique
PlaneNet) identifies oblique planes aligned with prostate apex and base, and
urethra.
Labels: All anatomical locations
and masks were manually labeled on the 3-plane scout volume, except prostate
and urethra were labeled on the high resolution sagittal T2 volume, and then
resampled on the 3plane scout volume.
Augmentation: All or part
of the augmentation techniques were implemented to improve model performance:
intensity augmentation, noise corruption, adaptive histogram equalization,
median filter, volume translation, rotation and cropping, z-score normalization
etc.
DL Methodology: DL segmentation networks were adapted for all the five models [5-7].
Accuracy Assessment: For CalLocNet model, dice score
was used as accuracy metric. For all ScanNets, accuracy was assessed by
calculating location from a segmentation mask and then the location error was
evaluated between GT and DL prediction. For predicted planes for prostate
apex/base and urethra, angle error between GT and DL-predicted planes was
calculated.RESULTS AND DISCUSSION
For CalLocNet, the
testing DICE score is 84.21%. For all ScanNets, the average location error is
less than 5mm, and angle error is less than 10 degrees. Fig. 1 shows an example
of predicted prostate mask vs ground truth prostate mask for the CalLocNet. Fig.
2 shows an example of predicted 7 locations from SagLocNet on the right
comparing with ground truth locations on the left. Table 1 presents the average
location error on the validation and testing datasets for the SagLocNet. Fig. 3
shows an example of predicted prostate apex/base plane and urethra plane
comparing to ground truth. Table 2 presents the average angle error on the
validation and testing datasets for the ObqliuePlaneNet.CONCLUSION
We introduced a fully
automated MR scanning workflow starting from a large FOV coil sensitivity
volume scan. Results indicate very high accuracy for volume prescription for
all prostate scans including 3plane scout scan. The generalized fully automated workflow
can be extended to other anatomies as well. Acknowledgements
No acknowledgement found.References
- Liu Y, Dawant BM. “Automatic localization of the anterior
commissure, posterior commissure, and midsagittal plane in MRI scans using
regression forests”. IEEE J Biomed Health Inform. 2015;19(4):1362-74.
-
Wieslaw L. Nowinski et.al., “Rapid and automatic calculation
of the midsagittal plane in magnetic resonance diffusion and perfusion images”,
Acad Radiol. 2006 May; 13(5): 652–663.
-
Bhushan, Chitresh & Shanbhag, Dattesh & Maximo, Andre
& Sreekumari, Arathi & Gui, Dawei & Patil, Uday & Pascual,
Brandon & Mullick, Rakesh & Yeo, Desmond & Foo, Thomas.
“Intelligent Knee MRI slice placement by adapting a generalized deep learning
framework”. Proceedings of ISMRM 2020
- K Manickam, M Shao, D Gui, C Bhushan, Dattesh D Shanbhag.
“Intelligent automatic slice prescription of scout scans of MSK MRI imaging
using surface coil sensitivities.” Proceedings of ISMRM 2023
- Deep Anand
et.al., “Realizing a robust multi-task
training strategy with deep learning: application in Spine MR image quality
assessment”, Proceedings of ISMRM 2023, p. 2950.
- Olaf Ronneberger, Philipp Fischer, Thomas Brox, “U-Net:
Convolutional Networks for Biomedical Image Segmentation”, Medical Image
Computing and Computer-Assisted Intervention (MICCAI), Springer, LNCS,
Vol.9351: 234--241, 2015