Automation of Pattern Recognition Analysis of Dynamic Contrast-Enhanced MRI Data to Assess the Tumor Microenvironment
SoHyun Han1, Radka Stoyanova2, Jason A. Koutcher3, HyungJoon Cho1, and Ellen Ackerstaff3

1Ulsan National Institute of Science and Technology, Ulsan, Korea, Republic of, 2Miller School of Medicine, University of Miami, Miami, FL, United States, 3Memorial Sloan Kettering Cancer Center, New York, NY, United States


Recently, a novel pattern recognition (PR) approach has been developed, identifying extent and spatial distribution of tumor microenvironments based on tumor vascularity. Here, our goal is to develop methods to minimize user intervention and errors from model-based approaches by introducing an automated algorithm for determining the number of classifiers. An SNR approach showed the highest accuracy at ~97% along five different tumor cell models with 104 slices total. The visualization of tumor heterogeneity (perfusion, hypoxia, necrosis) with automated analysis of DCE-MRI can reduce the need for manual expert intervention, extensive pharmacokinetic modeling, and could provide critical information for treatment planning.


The tumor microenvironment is heterogeneous, exhibiting severe structural and functional vascular abnormalities1, 2. Specifically, tumor hypoxia has been associated with aggressive tumor progression as well as resistance to radiation and/or chemotherapy, resulting in poor clinical outcome and prognosis1-5. Thus, visualization of the extent of hypoxia in tumors would provide critical information for strategic treatment planning. Conventionally, dynamic contrast enhanced (DCE)-MRI is used to investigate tumor perfusion by administration of Gd-DTPA with pharmacokinetic modeling. Recently, features of the tumor microenvironment were classified by pattern recognition approaches, such as Gaussian mixture model (GMM)6 or constrained non-negative matrix factorization (cNMF)7. These approaches minimize biases or errors from purely model-based fitting. However to-date, the choice of number of classifiers has been predefined manually. Here, we focus on developing and validating a robust algorithm for the automated determination of the number of classifiers, characterizing heterogeneity of tumor vascular features without subjective user intervention.


Different approaches of identifying the number of classifiers automatically were tested on DCE-MRI data from 5 different preclinical tumor models – HEK, LAPC-48, Myc-CaP9, PC-310, and RM-111 flank tumors in Nod/SCID mice. Experimental parameters for the DCE-MRI data acquisition: T1-weighted fast low-angle shot (FLASH) sequence with 3.2 ms echo time, minimum repetition time (TRmin), 256 repetitions (NR256, time points), 1 average, 15 mm x 15 mm FOV, 128x128 matrix, 1 mm slice thickness, and 5-7 slices to cover the entire tumor. For 5, 6, and 7 slices respectively, the resulting TRmin were 42.875 ms, 51.450 ms, and 60.025 ms with a corresponding temporal resolution of 5.487 s, 6.585 s, and 7.683 s. All animal experiments were performed according to protocols approved by the Institutional Animal Care and Use Committee of Memorial Sloan Kettering Cancer Center. As the principal component analysis (PCA, 1st analysis step) identifies the sources of largest variations, the number of classifiers (NC) was evaluated by three different criteria: signal enhancement, SNR, and half area under the curve (HAUC) of principal components (PCs). Signal enhancement was calculated as (max(PCs) – mean(PCs))/mean(PCs)×100 and a threshold for selecting significant PCs was set manually to 6000. SNR was calculated as max(PCs)/(2×std(PCs of pre contrast agent injection)) and a threshold set to 5 (SNRTh5). Additionally, a threshold of 2 (SNRTh2)was considered for tumors with low SNR. HAUC was calculated by adding the signal intensity during the initial half of total acquisition time (~10 mins) with a threshold set to 0.5×first HAUC. To assess the performance of the approaches to automatically determine NC, NC was also determined manually by an expert for each tumor slice. After selecting NC, cNMF was accordingly performed as described previously7. To generate pattern masks of tumor microenvironments, the representative curve showing the maximum value of normalized weights in a given pixel is allocated to each pattern mask. Also, pattern masks of pixels associated with a mixture of features are generated for pixels where the difference between maximum weight to other pattern weight is lower than 25%.


Three criteria – signal enhancement, SNR, and HAUC - for determining NC automatically were applied to five different preclinical tumor models with a total of 104 slices (HEK n = 6, LAPC-4 n = 6, MycCaP n = 2, PC-3 n = 2, RM-1 n = 4). The NCs obtained from these three criteria were compared to the NC determined by manual expert inspection and the corresponding accuracy is shown in Fig. 1. The accuracy was 75.00%, 86.54%, 97.12%, and 49.04% for enhancement, SNRTh5, SNRTh2 (considering necrotic tumor), and HAUC, respectively. The cNMF-derived representative pattern curves and corresponding weight maps are shown in Fig. 2. Based on vascular features characterizing specific tumor microenvironments7, red, green, and blue plots represent well-perfused, hypoxic, and necrotic tumor areas. Two different types of pattern masks characterizing the tumor slice were generated: “Decision map 1” was generated by finding maximum weight and allocating index as 1, 2, and 3 for necrotic, hypoxic, and well-perfused, respectively; “Decision map 2” was generated by adding the additional condition for mixtures. Of the three approaches tested, the SNR approach demonstrated to be the most useful strategy to automatically determine NC and could be optimized to account for DCE-MRI data with low contrast agent enhancement.


The visualization of tumor heterogeneity (perfusion, hypoxia, necrosis) with automated analysis of DCE-MRI can reduce the need for manual expert intervention, extensive pharmacokinetic modeling and could provide critical information for treatment planning.


We acknowledge support by NIH / NCI grants R01 CA163980 (RGB), R01 CA172846 (RGB, JAK), R24 CA083084 (SAI Core) and P30 CA008748 (Cancer Center Support Grant), the National Research Foundation of Korea (HJC) and Korean government grants No. 2010-0028684 (HJC) and No. 2014 R1A1A1 008255 (HJC).

The DCE-MRI data were acquired in collaboration between Dr. Blasberg’s (Dr. Moroz, technical support by Mr. Nisargbhai S. Shah) and Dr. Koutcher’s (Dr. Ackerstaff, technical assistance by Ms. Natalia Kruchevsky) laboratory and we gratefully acknowledge the permission to use these as testing data sets.


1. Vaupel P, Kallinowski F, Okunieff P. Blood flow, oxygen and nutrient supply, and metabolic microenvironment of human tumors: a review. Cancer Res. 1989; 49(23):6449-65.

2. Vaupel P. Tumor microenvironmental physiology and its implications for radiation oncology. Semin Radiat Oncol. 2004. Jul; 14(3):198-206.

3. Varlotto J, Stevenson MA. Anemia, tumor hypoxemia, and the cancer patient. Int J Radiat Oncol Biol Phys. 2005. Sep 1; 63(1):25-36.

4. Tatum JL, Kelloff GJ, Gillies RJ, Arbeit JM, Brown JM, Chao KS, Chapman JD, Eckelman WC, Fyles AW, Giaccia AJ, Hill RP, Koch CJ, Krishna MC, Krohn KA, Lewis JS, Mason RP, Melillo G, Padhani AR, Powis G, Rajendran JG, Reba R, Robinson SP, Semenza GL, Swartz HM, Vaupel P, Yang D, Croft B, Hoffman J, Liu G, Stone H, Sullivan D. Hypoxia: importance in tumor biology, noninvasive measurement by imaging, and value of its measurement in the management of cancer therapy. Int J Radiat Biol. 2006. Oct; 82(10):699-757.

5. Bache M, Kappler M, Said HM, Staab A, Vordermark D. Detection and specific targeting of hypoxic regions within solid tumors: current preclinical and clinical strategies. Curr Med Chem. 2008; 15(4):322-38.

6. Han SH, Ackerstaff E, Stoyanova R, Carlin S, Huang W, Koutcher JA, Kim JK, Cho G, Jang G, Cho H. Gaussian mixture model-based classification of dynamic contrast enhanced MRI data for identifying diverse tumor microenvironments: preliminary results. NMR Biomed. 2013. May; 26(5):519-32.

7. Stoyanova R, Huang K, Sandler K, Cho H, Carlin S, Zanzonico PB, Koutcher JA, Ackerstaff E. Mapping Tumor Hypoxia In Vivo Using Pattern Recognition of Dynamic Contrast-enhanced MRI Data. Transl Oncol. 2012. Dec; 5(6):437-47.

8. Craft N, Shostak Y, Carey M, Sawyers CL. A mechanism for hormone-independent prostate cancer through modulation of androgen receptor signaling by the HER-2/neu tyrosine kinase. Nat Med. 1999. Mar; 5(3):280-5.

9. Watson PA, Ellwood-Yen K, King JC, Wongvipat J, Lebeau MM, Sawyers CL. Context-dependent hormone-refractory progression revealed through characterization of a novel murine prostate cancer cell line. Cancer Res. 2005. Dec 15; 65(24):11565-71.

10. Webber MM, Bello D, Quader S. Immortalized and tumorigenic adult human prostatic epithelial cell lines: characteristics and applications. Part I. Cell markers and immortalized nontumorigenic cell lines. The Prostate. 1996. Dec; 29(6):386-94.

11. Baley PA, Yoshida K, Qian W, Sehgal I, Thompson TC. Progression to androgen insensitivity in a novel in vitro mouse model for prostate cancer. J Steroid Biochem Mol Biol. 1995. May; 52(5):403-13.


Figure 1: Accuracy of different methods to automatically determine the number of classifiers NC, characterizing the number of tumor microenvironments based on tumor vascular features.

Figure 2: For a representative tumors slice, (left) three cNMF curves, characteristic for contrast agent uptake behavior in well-perfused (red), hypoxic (green), and necrotic (blue) tumor areas, (center) their corresponding weighted maps, and (right) resulting pattern masks, assigning each tumor pixels to a specific environment or mixture thereof.

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)