Mahdi Orooji1, Mehdi Alilou2, Rachel Sparks3, Mirabela Rusu4, Nicolas Bloch5, Ernest Feleppa6, Dean Barratt7, Lee Ponsky8, and Anant Madabhushi2
1Biomedical Engineering, CenteCase Western Reserve University, Cleveland, OH, United States, 2Biomedical Engineering, Case Western Reserve University, Cleveland, OH, United States, 3Centre for Medical Image Computing, University College of London, London, United Kingdom, 4Albany, NY, United States, 5Boston Medical Center, Boston, MA, United States, 6Lizzi Center for Biomedical Engineering, Riverside Research, New York, NY, United States, 7University College London, London, United Kingdom, 8University Hospital Case Medical Center, Cleveland, OH, United States
Synopsis
To evaluate whether the combination of
computer extracted or radiomic image parameters from two complementary
modalities, MRI-TRUS can enable better prediction of presence of prostate
cancer compared to either modality individually. We considered 12 slides who
underwent MRI, TRUS prior to radical prostatectomy. Deformable co-registration
methods were used for spatially aligning the pre-operative in vivo MRI and ultrasound with the ex vivo whole mount radical prostatectomy specimens to establish
the ground truth for cancer extent on the imaging. It yielded the best
separability between cancer and non-cancer regions with an Area under the
operating characteristic curve of 0.88.PURPOSE
Recently there has
been a great deal of interest in developing computer aided diagnosis systems
for identifying prostate cancer presence in vivo on MRI and ultrasound
separately1-3, no work we are aware of has attempted to address the
issue of fusing computer derived features from MRI and ultrasound to create the
best possible predictor of cancer in vivo. In this work we attempt a systematic
and quantitative evaluation of the discriminability of computer extracted MRI
and ultrasound features in terms of cancer detection in patients undergoing
radical prostatectomy.
METHODS
Our study design comprised 12 2D planar images
obtained from the MRI and US scans of 3 patients, all of whom had biopsy
confirmed prostate cancer and scheduled for a radical prostatectomy. A 3D
B-mode ultrasound scan was performed followed by a 3 Tesla MRI prior to
surgery. Following surgery and histologic sectioning of the gland via a
microtome, the H&E stained whole mount histologic (WMH) sections were
digitized via a whole slide scanner and the regions of cancer annotated by an
expert pathologist. Deformable co-registration methods were used to spatially
align the in vivo MRI, TRUS, and ex vivo histology. In particular, we
used fully automatic Multiattribute probabilistic prostate elastic registration
(MAPPER) approach to fusion of ultrasound and MRI4. We also manually
delineated
corresponding landmarks between MRI and WMH for deformable
co-registration of WMH to MRI. A total of 129 computer extracted image features
including Haralick, Gabor, Law, LBP, Laplacian features were extracted from
both the prostate MRI and TRUS. Each of the computer extracted MRI and
ultrasound features were then ranked via the Fisher criteria to identify the
features that best identified the region of cancer. Figure 1 illustrates the
MRI-TRUS-WMH registration and mapping of the cancer extent on MRI and TRUS.
RESULTS AND
DISCUSSION
The top 3 features
for each modality and corresponding Fisher criteria values are shown in
Table.1. The classification is per region of interest (ROI), i.e. the texture
features for the cancerous part of the prostate is compared to the texture
features of the noncancerous confounding regions. Top three texture features,
contrast variance, contrast entropy, and contrast inverse moment were selected by
the theoretical linear discriminant analysis (LDA) classifier for MRI and
yielded an area under the receiver operating characteristic
curve (AUC) of 0.83, 0.77 and 0.70 for identifying cancerous
ROIs in MRI. By comparison, top three most predictive features identified for
TRUS were contrast inverse moment, contrast variance, and contrast entropy.
These features yielded an AUC of 0.75, 0.69, and 0.66, respectively. By
combining the top two texture features on MRI and the most informative texture
feature on TRUS, the LDA based predictor yielded an AUC of 0.88 in predicting
presence of prostate cancer. Figure 2 illustrates the scatter plot of the
prostate cancer versus the non-cancer cases in three dimensional most
informative texture feature space.
CONCLUSION
We presented a
framework to rank the performance of computer extracted MRI and ultrasound
features in terms of their ability to identify prostate cancer. Our results in
a small cohort suggests that we may be
able to combine the MRI and ultrasound radiomic features to create a better
classifier for prostate cancer detection compared to MRI or ultrasound alone.
Acknowledgements
No acknowledgement found.References
[1] Emilie Niaf, Olivier Rouvire, Florence Mge-Lechevallier,
Flavie Bratan, Carole Lartizien,” Computer-aided diagnosis of prostate cancer
in the peripheral zone using multiparametric MRI”, Physics in Medicine and
Biology 2012.
[2]Tiwari P, Kurhanewicz J, Rosen M, Madabhushi A.
(2010) Semi supervised multi kernel (SeSMiK) graph embedding: identifying
aggressive prostate cancer via magnetic resonance imaging and spectroscopy. Med
Image Comput Comput Assist Interv. 13(Pt 3):666-73
[3]Mehdi Moradi, Parvin Mousavi, Purang
Abolmaesumi, Computer-Aided Diagnosis of Prostate Cancer with Emphasis on
Ultrasound-Based Approaches: A Review, Ultrasound in Medicine & Biology,
Volume 33, Issue 7, July 2007, Pages 1010-1028, ISSN 0301-5629
[4] Sparks, Rachel and Nicolas Bloch, B. and
Feleppa, Ernest and Barratt, Dean and Moses, Daniel and Ponsky, Lee and
Madabhushi, Anant, “Multiattribute probabilistic prostate elastic registration
(MAPPER): Application to fusion of ultrasound and magnetic resonance imaging”,
Medical Physics, 42, 1153-1163 (2015),