Samuel Frankel1, Ananya Panda2, Debra McGivney2, Gregory O'Connor1, Alice Yu3, Mark Griswold2,4,5, Chaitra Badve1,2,4, and Vikas Gulani2,4,5
1School of Medicine, Case Western Reserve University, Cleveland, OH, United States, 2Radiology, Case Western Reserve University, Cleveland, OH, United States, 3Johns Hopkins University, Baltimore, MD, United States, 4Radiology, UH Cleveland Medical Center, Cleveland, OH, United States, 5BME, Case Western Reserve University, Cleveland, OH, United States
Synopsis
Prostate cancer and prostatitis can have considerable
overlap on conventional MR imaging.
Texture analysis on multiparametric MRI shows promise in
characterization of prostate, but has not been used on quantitative prostate
maps. Here we utilize texture analysis
on magnetic resonance fingerprinting (MRF) maps of prostate for
characterization of prostate lesions. Results
show that texture features can differentiate cancer and non-cancerous
transition zone and between grades of cancer in peripheral zone. This could add
value to MRF-based relaxometry and conventional MRI to improve lesion
characterization.
Introduction
Previous literature describes
patterns of pixel attributes, or “texture analysis,” to identify prostate
cancer on qualitative T1-weighted and T2-weighted
magnetic resonance images, and on ADC maps1-3. Second-order statistical features, or “texture
features” characterize pairs of pixels in an image, providing statistical information
undetectable to the naked eye. To date,
texture analysis has not been used on prostate tissue relaxometry maps. Here we
apply this to Magnetic Resonance Fingerprinting (MRF) T1 and T2
relaxation times4. Previous MRF analysis of prostate has shown that the
combination of T1, T2 and ADC values can differentiate between
normal peripheral zone (PZ) and PZ prostate cancer and between different
Gleason grades (GS)5. However, some residual overlap was still
present, in particular between prostatitis and cancer. The purpose of this work
is to apply texture analysis to improve upon quantitative MRF prostate data and
further characterize prostatic pathologies.Methods
In this IRB and HIPPA compliant
study, 74 patients with clinical suspicion of prostate cancer underwent
MRI (High resolution T2 w, diffusion weighted imaging using echo
planar imaging with b-values 50-1400 s/mm2) and MRF-FISP acquisition
as previously described5 (resolution 1 x 1 x 5 mm3) on a
3T scanner. Regions of Interest (ROIs) (mean size 66.3 mm2, 6-445 mm2)
were drawn on MRF maps in both cancer suspicious regions and normal tissue by
a radiologist (7 years’ experience) based on clinical MRI reads by another
radiologist (16 years experience) using a MATLAB-based GUI (Fig. 1). 58 patients with 65 PZ lesions and 24 patients
with 24 transition zone (TZ) lesions underwent targeted biopsy based on MRI
findings. 26 second order texture
features were calculated in MATLAB for each ROI using Haralick texture features6
derived from Gray Level Co-occurrence Matrices (GLCM) and Gray Level Run Length
Matrices (GLRLM). Spearman rank
correlation coefficients were used to remove redundant texture features. The remaining features were correlated with
the gold standard pathologic diagnosis.
Student’s t-tests were performed on texture features between different
diagnoses (GS 6-9 cancer and prostatitis) to characterize the lesions. Bonferroni correction was applied to correct
for multiple comparisons. For
significant texture features, logistic regression models were created to
calculate Receiver Operating Characteristic (ROC) curves, and Area Under the
Curve’s (AUC) were calculated.Results
Pathologic diagnosis for lesion
ROIs were: GS 6 / low grade cancer (PZ:n=8;TZ:n=3), GS 7/ intermediate grade cancer
(PZ:n=33;TZ:n=9), GS ≥8 / high grade cancer (PZ:n=11) and prostatitis (PZ:n=13;TZ:n=12). Spearman Correlation Coefficients were
calculated, resulting in the removal of 8 texture features (18 remaining). Results for texture analysis in PZ and TZ are
presented in Table 1. After Bonferroni
corrections, T2 cluster shade was significantly different between
intermediate and high grade cancer for PZ lesions (p=0.01;AUC=0.67). For transition zone, T1 energy and
entropy were significantly different between cancer and prostatitis (p=0.011,p=0.004;AUC=0.80,AUC=0.81
respectively) (Fig. 2, 3). T2 informational measure of correlation 1
(IMC1) and inverse difference were significantly different between normal TZ
and cancer (p=0.002,p=0.016;AUC=0.86,AUC=0.74 respectively) (Fig. 2, 4). T1 energy and entropy and
T2 IMC1 and inverse difference were significantly different between
TZ cancer and non-cancer tissue (combined normal TZ and prostatitis) (p=0.039,p=0.020,p=0.003,p=0.048;
AUC=0.72,AUC=0.71,AUC=0.83,AUC=0.68 respectively).Discussion
This work shows application of
texture analysis to MRF maps. As texture features are calculated based on local
voxel intensity relationships, they can reflect local spatial variation
patterns in pathologic tissue versus normal tissue and improve tissue characterization.
In this work, MRF- based texture analysis could provide additional separation
of intermediate and high grade cancers in PZ compared to MRF relaxometry alone5.
In TZ, multiple texture features showed significance, similar to the
results reported by previous studies done on qualitative T-2 w
images1-3. However using MRF, we also
obtained additional T1 texture features for differentiation. Some of
these second order statistical features have straightforward correlates, e.g. entropy
measures pairwise pixel randomness similar to heterogeneity, and IMC1 is a
function of entropy, while others lack a visible correlate. Clinically, TZ cancers are known to appear
more homogenous than TZ non-cancerous tissue.
Texture changes may indicate inherent tissue differences between lesion
and non-lesion tissue. Texture analysis
using MRF maps can potentially be combined with radiomics and pathonomics to
improve lesion detection and characterization.
One limitation of this study is that we used targeted biopsy ROIs for
texture analysis, whereas whole mount prostatectomy specimens can give a more
accurate correlation of texture features with pathology. Conclusion
This work shows the feasibility
and utility of applying texture analysis methods directly to quantitative MRF
maps in differentiating grades of cancer in peripheral zone and normal tissue
versus cancers in transitional zone of prostate. Acknowledgements
This work was made possible with support from NIH grants 1R01EB016728, 1R01BB017219, 1R01DK098503, 1R01CA208236, as well as Siemens Healthineers.References
1.
Bates, A. & Miles, K. Prostate-specific membrane antigen PET/MRI
validation of MR textural analysis for detection of transition zone prostate
cancer. European Radiology, 2017;Epub.
2.
Khalvati, F. et al., Automated prostate cancer detection via
comprehensive multi-parametric magnetic resonance imaging texture feature
models. BMC Medical Imaging 2015;15:27.
3.
Nketiah, G., et al. T2-weighted MRI-derived textural features reflect
prostate cancer aggressiveness:
preliminary results. European Radiology, 2017;27(7):3050–3059.
4.
Ma, D., et al. Magnetic resonance
fingerprinting. Nature, 2013;495(7440):187–192.
5. Yu, A. C., et al. Development
of a combined MR fingerprinting and diffusion examination for prostate
cancer. Radiology 2017;283(3):729–738.
6.
Haralick, R. Textural features
for image classification. IEEE
Transactions on Systems, Man and Cybernetics, 1973:6;610-621