Final Report Summary - 31P_SPECTRA_3T (Phosphorus MR Spectroscopic Imaging of Brain Tumors at 3T) Phosphorus magnetic resonance spectroscopic imaging (31P MRSI) is a kind of non-invasive MR spectroscopic imaging technique that detects the phosphorus instead of proton containing metabolites of the brain. 31P MRS provides in-vivo quantitative information about the energy metabolism, oxygen state and pH within a given region of interest. Phosphocreatine (PCr), phosphorylcholine (PC), phosphorylethanolamine (PE), inorganic phosphate (Pi), glycerophosphorylcholine (GPC), glycerophosphorylethanolamine (GPE), and three different peaks for the ATP molecule, g-ATP, a-ATP, and b-ATP peaks, are the major metabolites observed with 31P MRS. This project was motivated by the fact that, despite the advantages of 31P MRSI, phosphorus is 15 times less abundant in the body than proton molecules, and therefore phosphorus spectroscopy has not been widely used for clinical applications at lower field strengths. The wider availability of high field scanners and multi-channel radiofrequency (RF) surface coils have increased the sensitivity and accuracy of MR imaging and phosphorus MR spectroscopic imaging of brain tumors through higher signal-to-noise ratio (SNR) and improved spectral resolution. The goal of this project was to apply phosphorus magnetic resonance spectroscopic imaging accurately at high field 3T scanners to add new information regarding the characteristics of brain tumors and produce new metrices to estimate the aggresiveness of a brain tumor using 31P MRSI peak intensities. The sub-objectives of the project were,- to estimate T1 relaxation times of phosphorus containing metabolites within the human brain at 3T,- to compare different spectral data analysis strategies to define the most accurate way of quantifying phosphorus spectral data acquired from the human brain at 3T,- to analyze the spatial heterogeneity and characteristics of brain tumors using 31P MRSI at 3T, and- to define new metrices that can characterize how aggressive a brain tumor is based on the intensities of phosphorus containing metabolites.All four aims of the project were accomplished with minor deviations. The clinical scan time limitations require a short repetition time (TR) for 31P MRSI that results in signal saturation, which might cause an incorrect estimation of metabolite levels and ratios. In the first study of the project, we measured T1 relaxation times, and also estimated the partial saturation factors of phosphorus containing metabolites of the human brain at 3T to correct for saturation effect in spectral quantification. The saturation factors for metabolite ratios were calculated using the estimated T1 values for each metabolite and the repetition time. The mean (±std) T1 values of PCr, g-ATP, a-ATP, b-ATP, GPC, GPE, Pi, PC, and PE peaks were estimated as 4.61 (±0.53) 1.88 (±0.72) 2.08 (±0.53) 2.13 (±0.66) 2.13 (±0.64) 4.78 (±0), 3.58 (±0.75) 3.26 (±1.37) 3.75 (±0.81) seconds, respectively. The range of mean saturation factors was 0.76 to 1.40 for the ratios of interest. Low SNR, baseline distortions, quantification errors and fitting errors all play a role in the accuracy of T1 measurements. In our study, to minimize baseline distortion effect, the data were quantified in AMARES and estimated peak values were fit into a theoretical model function in MATLAB. A priori information of relative frequency differences of peaks was also provided to accurately find the 31P MRS peaks in AMARES. As a result, R2 values were higher than 0.95 for all the fits of individual peaks. GPE and GPC, and PC and PE peaks were not well separable for all the subjects. T1 of GPE peak was measurable in only one subject. T1 measurement of phosphorus containing metabolites in brain tumor patients was not achieved due to long scan time, which was almost one and a half hours for each subject including the preparation, the scout scan, shimming and 44 minutes of 31P MRS scan time. The second objective of this project was to compare time domain fitting and frequency domain analysis for accurate quantification of 31P MRSI data of human brain at 3T. Phosphorus is 15 times less MR sensitive than proton, and low signal-to-noise ratio (SNR) makes 31P MRSI quantification quite challenging. For this part of the project, Dr. Ozturk-Isik collaborated with Dr. Jason Crane from UCSF Surbeck Laboratory of Advanced Imaging, who is one of the developers of SIVIC platform. The amplitude, damping, and frequency factors were calculated for each peak with AMARES. SIVIC program was used for frequency domain analysis. The spectra were converted into a SIVIC readable format in MATLAB. After Fourier transform, the area under each peak were estimated in frequency domain in SIVIC. The area under each peak for SIVIC or amplitude for AMARES was normalized with that of PCr. A Bland Altman statistical test, which plots the difference against the mean of two observations, was used to test the agreement of SIVIC and AMARES methods. The peak ratio estimates of AMARES and SIVIC were very similar according to the Bland Altman test results. AMARES was observed to provide better peak estimates for more noisy spectra. The third objective of this project was to analyze the spatial heterogeneity and characteristics of brain tumors using 31P MRSI at 3T. Several studies have reported 31P-MRSI peak differences between brain tumors and healthy tissue. For this part of the study, three healthy volunteers and 11 brain tumor patients, who provided informed consent in accordance with the Ethics Review Board regulations of Yeditepe University, were scanned on a 3T clinical MR scanner (Philips Medical Systems, Best, Netherlands), equipped with a dual channel 31P/1H quadrature head coil. The peak heights and ratios of the metabolites were calculated for each voxel. Logistic regression, and support vector machine with a linear and polynomial kernel were used to classify tumor voxels based on their phosphorus MR spectroscopic imaging characteristics, which has not been previously reported in the literature. Recursive feature elimination (RFE) was used to determine the most effective features for classification. Four-fold cross validation was used to measure the relative performances of the three machine learning methods. GPC, PE, g-ATP, GPC+GPE, PC+PE, PC/GPC, and PE/GPE, were determined as the most effective features for classification. Tumor voxels displayed a higher GPC/PCr, PE/PCr, g-ATP/PCr, (GPC+GPE)/PCr, (PC+PE) /PCr, and PE/GPE than normal voxels (p<0.004). Support vector machine with a linear kernel, support vector machine with a polynomial kernel, and logistic regression were able to distinguish normal voxels from tumor voxels based on their phosphorus MR spectroscopic peak parameters, with 80.74%, 77.69%, and 90.51% sensitivities, and 68.37 70.20 and 72.56 specificities, respectively. The resultant accuracies were 74.80 73.84 and 81.71 for the support vector machine with a linear kernel, support vector machine with a polynomial kernel, and logistic regression methods, respectively. The results of this study showed that support vector machine and logistic regression could be employed for defining classification models to discriminate brain tumor from normal tissue based on 31P-MRSI data at 3T. Logistic regression resulted in a higher sensitivity, specificity and accuracy than both SVM methods. The main limitation of this study was small subject population resulting in rather low accuracy. The fourth objective of the grant was to define metrices that can characterize how aggressive a brain tumor is based on the intensities of phosphorus containing metabolites. A similar study has been performed for 1H-MRSI, but no such metrices have been studied for phosphorus MR spectroscopic imaging yet. For this part of the study, thirteen brain tumor patients’ and five volunteers’ 31P-MRSI data were used. The spectra were processed using AMARES within jMRUI, and MATLAB was used to estimate metabolite peak heights and metabolite ratios. A T1 weighted fast field echo (FFE) MR image was acquired, and tumor and healthy voxels were determined out of this anatomical image. For each subject, a linear regression line with a zero intercept was fit to the voxel intensities for a given metabolite ratio. A z-score was computed for the distance between the levels of the ratio for each voxel and the linear regression line fit. The voxels that are more than two standard deviations away from the regression line were discarded, and a new fit was performed with the remaining points until no outliers remain. The final regression line was the estimate of the ratio that can be seen in normal voxels. The final z-scores for each voxel were determined as the distance from the final regression line minus the mean value of the distances of healthy voxels, divided by the standard deviation of the distances of healthy voxels. This procedure was performed for each metabolite peak ratio and the resultant metric was named after the corresponding ratio, like Pi/PCr index. Mainly, the metabolic indices that were indicated as significantly different between tumor and healthy tissue by the third study were analyzed. GPE/PCr, GPC/PCr, PC/PCr, PE/PCr, g-ATP/PCr, Pi/PCr indices were higher in the tumor voxels of the brain tumor patients than volunteers. The definition of quantitative metrices based on 31P-MRSI metabolic indices might provide the clinicians with a tool to identify the most aggressive parts of a lesion that can be deliniated as a target for therapy. A website for this project has been prepared and routinely updated, and the address of the website is http://esinozturkisik.com/31P_SPECTRA_3T.html. A project logo was designed, which can be found attached to this final report and at the website. The project abstract, objectives, publications, some answers to the frequently asked questions about the project, dissemination activities, and information about the Medical Imaging Laboratory that was founded by the support of this project at Yeditepe University can be located at the project website. An additional grant was obtained from The Scientific and Technological Research Council of Turkey (TUBITAK) for fast phosphorus spectroscopic imaging using compressed sensing technique, and this project has been conducted in close collaboration with the researchers at Uludag University Hospital. This additional project started on October 1, 2012, and the period of this grant is two years. The researcher established a research agreement between Philips Healthcare and Uludag University. Through this agreement, the researcher has implemented compressed sensing accelerated 31P MRSI technique and has installed it on the 3T clinical MR scanner at Uludag University Hospital. A link was added to the website for a brief description of this project that can be accessed by pressing the ‘Other Grants’ button. In conclusion, as a result of this Marie Curie IRG grant, Dr. Ozturk-Isik has successfully established her integration to ERA. She has founded Medical Imaging Laboratory at Yeditepe University, and initiated several collaborations. She has supervised one Ph.D. student, three Master’s students, and 31 undergraduate students for their research projects. She has also published several conference papers and articles in the field of phosphorus MR spectroscopic imaging, and other MR imaging techniques. She has also disseminated her knowledge through invited talks and lectures. So far, Dr. Ozturk-Isik has developed and taught eight different undergraduate and graduate level courses. She has also continued closely collaborating with medical doctors at Yeditepe University Hospital, Uludag University Hospital, and American Hospital. As a result of this project, 31P MRSI has been more widely used, and would provide important metabolic information for diagnosis, treatment planning and follow up of brain tumors.