Shenglan Chen1, Yinxia Zhao2, Xintao Zhang2, Tianyun Zhao1, Mario Serrano-Sosa1, Xiaodong Zhang2, and Chuan Huang1,3,4
1Biomedical Engineering, Stony Brook University, Stony Brook, NY, United States, 2Radiology, The Third Affiliated Hospital of Southern Medical University (Orthopaedic Hospital of Guangdong Province), Guangzhou, China, 3Radiology, Stony Brook Medicine, Stony Brook, NY, United States, 4Psychiatry, Stony Brook Medicine, Stony Brook, NY, United States
Synopsis
Bone marrow fat fraction (BMFF) has been recognized as one
of the quantitative image biomarkers to identify abnormal bone density using
modified Dixon sequence. However, this method requires manual segmentation
which limits its adoption in clinical practice. In this study, we developed a
fully automated radiomics pipeline using deep learning based segmentation and
validated its performance comparable to manual segmentation. This finding will
facilitate the clinical utility of the entire pipeline as a screening tool for
early detection of abnormal bone density.
Introduction
Osteoporosis
has become a major public health problem worldwide; after 50 years of age, the risk
of osteoporotic fractures is 50% for women and 20% for men among many Western
populations[1]. Therefore,
it is critically important to identify patients at high risk of osteoporosis at
an early stage. A previous study[2] demonstrated the capability of bone
marrow fat fraction (BMFF) extracted using Modified Dixon (mDixon) Quant to
predict abnormal bone density, however, the clinical utility is limited by the requirement
of manual segmentation, and the predictive power of more quantitative imaging
biomarkers remains to be explored. In this study, we demonstrate a fully
automated end-to-end radiomics pipeline using reliable segmentation from uNet[3] convolutional neural network (CNN)
architecture. Materials and Methods
Population and image acquisition: A total
of 257 participants without known spinal tumor, history of trauma, dysplasia,
spinal surgery or hormone therapy were recruited in this prospective study with
approval by local IRB. All patients underwent QCT examinations to obtain bone
mineral density (BMD) of lumbar vertebrae (L1 to L3). The recommended threshold
of 120 mg/cm3 by the International Society for Clinical Densitometry
(ISCD)[4] was used
to determine abnormal bone density. mDixon Quant with six-echo, seven fat peaks
modeling was obtained from a 3.0T MR scanner (Ingenia, Phillips, Amsterdam,
Netherlands) and T2* correction was then performed to quantify the vertebral
fat fraction.
Image processing and analysis: Rectangular
regions-of-interest (ROIs) were delineated in mid-sagittal view of each
vertebral body on the fat fraction maps by a radiologist with 12 years of
experience. The dataset was temporally divided into two sets (Training set n =
142 with 74 in abnormal bone density class, and test set n = 64 with 26 in
abnormal bone density class) to develop radiomics signature. uNet CNN was developed
on the same training set to segment L1-L3 from fat fraction maps. A total of 92
radiomic features were extracted using in-house Matlab software[5], including
first-order, GLCM and GLRLM features from both fat fraction and T2* maps.
Feature pre-selection was performed to identify features with high
reproducibility and predictive power (Intraclass correlation coefficient >
0.5, Mann Whitney U-test p<0.05, Spearman
correlation |ρ|<0.8). The final prediction model was
built via least absolute shrinkage and selection operator (LASSO)
and logistic regression with 3-fold
cross-validation. The receiver operating characteristic
(ROC) analysis was performed to evaluate the model performance on the test
dataset with ROIs drawn by the radiologist and uNet. Results and Discussion
As
shown in Figure 1, uNet achieved a mean Dice coefficient of 0.912±0.062 in the
test set. In terms of the performance in predicting abnormal bone density, the
CNN segmentation achieved AUC (0.921) comparable to manual segmentation (0.931)
for the radiomics-based prediction model. This radiomic model achieved a
superior performance compared to previously model based only on BMFF and
clinical variables [2]. In particular, the
model achieved excellent NPV at 0.914/0.911 for manual and uNet segmentation
respectively, which is clinical important to identify patients with normal bone
density as a screening tool. Receiver operating characteristic data are shown
in Figure 2 and prediction metrics are summarized in Table 1. Conclusion
A fully
automated radiomics pipeline using reliable segmentation from CNN architecture
was developed and validated with comparable performance to manual segmentation.
This finding will help advance the clinical adoption of the entire pipeline as a screening tool for early
detection of abnormal bone density. Acknowledgements
No acknowledgement found.References
1. Liu, J., et al., State of the art in osteoporosis risk assessment and treatment. J
Endocrinol Invest, 2019. 42(10): p.
1149-1164.
2. Zhao,
Y., et al., Prediction of Abnormal Bone
Density and Osteoporosis From Lumbar Spine MR Using Modified Dixon Quant in 257
Subjects With Quantitative Computed Tomography as Reference. Journal of
Magnetic Resonance Imaging, 2019. 49(2):
p. 390-399.
3. Ronneberger,
O., P. Fischer, and T. Brox. U-Net:
Convolutional Networks for Biomedical Image Segmentation. 2015. Cham:
Springer International Publishing.
4. Engelke,
K., et al., Clinical use of quantitative
computed tomography and peripheral quantitative computed tomography in the
management of osteoporosis in adults: the 2007 ISCD Official Positions. J
Clin Densitom, 2008. 11(1): p.
123-62.
5. Renee
Cattell, S.C., Chuan Huang, Robustness of
Radiomic Features in MRI: Review and A Phantom Study. Visual Computing for
Industry, Biomedicine, and Art, 2019.