Amy L. Kaczmarowski1, Kenneth Iczkowski2, William A. Hall3, Ahmad M. El-Arabi4, Kenneth Jacobsohn4, Paul Knechtges1, Mark Hohenwalter1, William See4, and Peter S. LaViolette1
1Radiology, Medical College of Wisconsin, Milwaukee, WI, United States, 2Pathology, Medical College of Wisconsin, Milwaukee, WI, United States, 3Department of Radiation Oncology, Medical College of Wisconsin, Milwaukee, WI, United States, 4Urology, Medical College of Wisconsin, Milwaukee, WI, United States
Synopsis
Radiological-pathological
correlation is being used to validate prostate cancer imaging technology. This
study combines these two modalities with machine learning to generate
predictive maps of histological features (i.e. new contrasts) based on
segmented histology. We find that epithelium density maps highlight regions
pathologically confirmed as Gleason grade ≥3. This allowed the prediction of prostate
cancer presence based solely on non-invasive imaging in 23 of 26 cases.Purpose
Radiological-pathological (Rad-Path) correlation has
recently allowed the validation of prostate cancer imaging technology
1,2. Multiparametric MRI (MP-MRI)
has shown to be highly accurate at locating targets for biopsy
3. However, difficulties still remain in this
method of prostate cancer detection and grading
4,5. We present a new technique that combines the two
cross-modal datasets of MP-MRI and the gold standard, pathological grading, with
machine learning to generate predictive maps of histological features. These
maps can be used to predict prostate cancer presence with non-invasive imaging
alone.
Methods
Between October of 2014 and September of 2015, twenty-six patients undergoing prostatectomy were imaged
<2 weeks pre-op using a 3T MRI machine (GE) and MP-MRI. This included
diffusion imaging with 10 b-values, dynamic contrast enhanced (DCE) and T2
weighted imaging. We designed and used custom 3D-printed patient-specific slicing
molds to section the prostates in the same orientation as the imaging was
acquired
6,7. Whole-mount sections were paraffin-embedded, H&E stained,
digitized, and contoured using custom code developed in Matlab (Mathworks Inc).
Segmentation features included lumen, stroma, and epithelium/nuclei (Figure 1). Digitized
histology was co-registered to the T2-weighted MRI using Matlab software and non-linear
warping (Figure 2). A partial least squares regression (PLS) machine-learning algorithm
was trained with co-localized MRI voxel values and corresponding histology
features. A leave one out approach was used to generate predictive maps of each
of the three histological features (lumen, stroma, and epithelium/nuclei) for
each patient using the histology and corresponding MRI values from all other
patients (Figure 3). Maps of each feature were then compared to Gleason (G) graded whole
mount histology, performed by a board-certified pathologist.
Results
All 26 patients had
benign/atrophic glands and areas of G3 cancer. Twenty patients had regions of G4
and three patients had G5 nodules. Quantitative MRI-derived PLS trained maps of epithelium/nuclei were more
predictive of the location of cancer than those of stroma and lumen (Figure 4). In 6 of 6
patients with G3 alone, the epithelium density maps indicated high percentage
regions that co-localized with the pathologically confirmed G3 cancer.
Likewise, in 17 of 20 G4 cases and all three patients with G5 cancer, abnormally
high epithelium percentages co-localized with the pathologically confirmed
regions (Figure 4).
Conclusion
We generated machine learning trained maps of histological features
based on non-invasive quantitative MR imaging. We found that locations with high
epithelium/nuclei density indicated regions of G3+ cancer. More research is
necessary to determine how this new machine learning based biomarker should be
used for diagnostic purposes.
Acknowledgements
Advancing a Healthier Wisconsin
State of Wisconsin Tax Check Off Program for Prostate Cancer Research
NCI U01-CA176110-01A1
References
1. Nagarajan R, Margolis D, Raman S, et
al. Correlation of Gleason scores with diffusion-weighted imaging findings of
prostate cancer. Advances in urology. 2012;2012:374805.
2. Mazaheri
Y, Hricak H, Fine SW, et al. Prostate tumor volume measurement with combined
T2-weighted imaging and diffusion-weighted MR: correlation with pathologic
tumor volume. Radiology. 2009;252(2):449-457.
3. Panebianco
V, Barchetti F, Sciarra A, et al. Multiparametric magnetic resonance imaging
vs. standard care in men being evaluated for prostate cancer: a randomized
study. Urologic oncology. 2015;33(1):17
e11-17.
4. Panebianco
V, Barchetti F, Barentsz J, et al. Pitfalls in Interpreting mp-MRI of the
Prostate: A Pictorial Review with Pathologic Correlation. Insights into imaging. 2015.
5. Notley
M, Yu J, Fulcher AS, Turner MA, Cockrell CH, Nguyen D. Diagnosis of recurrent
prostate cancer and its mimics at multiparametric prostate MRI. Br J Radiol. 2015;88(1054):20150362.
6. Priester
A, Natarajan S, Le JD, et al. A system for evaluating magnetic resonance
imaging of prostate cancer using patient-specific 3D printed molds. American journal of clinical and
experimental urology. 2014;2(2):127-135.
7. Turkbey
B, Mani H, Shah V, et al. Multiparametric 3T prostate magnetic resonance
imaging to detect cancer: histopathological correlation using prostatectomy
specimens processed in customized magnetic resonance imaging based molds. The Journal of urology. 2011;186(5):1818-1824.