Magnetic resonance fingerprinting¹ (MRF) has been shown to be an efficient and accurate method to produce simultaneous quantitative maps of tissue properties such as T₁ and T₂ relaxation times. These quantitative data can aid in developing objective algorithms for assessing disease by imaging. Previous works have looked to draw conclusions from quantitative MRF data in prostate cancer² and brain tumors³, however, interpretation of this multidimensional data is difficult. In this work, we consider additional mathematical tools for interpreting the multidimensional quantitative MRF data for differentiation of normal and diseased states within the prostate. More specifically, we consider if the tissue parameters of T₁ and T₂ relaxation times acquired from MRF combined with the apparent diffusion coefficient (ADC) maps from traditional echoplanar imaging (EPI) acquisitions can delineate between prostate cancer, prostatitis, and normal tissue within the peripheral zone. Our aim is to find a simple score that can be quickly and widely used to aid in diagnosis.

This study is HIPAA and IRB compliant. Patients were scanned on 3T scanners (Siemens Verio and Skyra) using an MRF FISP⁴ acquisition. Analysis on the MRF maps was provided by two radiologists, each drawing regions of interest in both suspicious lesions as well as in the normal peripheral zone (NPZ). We analyze the data from one radiologist. The T₁, T₂, and ADC values were averaged in each of these locations and diagnosis of the lesions was made using targeted biopsy in 22 of the cases and the remaining with untargeted transrectal ultrasound biopsies. Biopsy specimens were read clinically by a genitourinary specialized pathologist. Gleason score was assigned. Gleason 3+3=6 cancers were considered low grade, 3+4=7 intermediate, and 4+3=7 and higher were considered high grade cancers.The data were gathered into a matrix and standardized so that each column had standard deviation of one, but was not centered. Principal component analysis (PCA) was applied to the three-dimensional data set to project each point into one-dimensional space using the right singular vector associated with the largest singular value of the matrix X.The one-dimensional projection of the data using principal component analysis is shown in Figure 1. The results have been shifted so that a delineation can be visualized at zero; data points to the right of zero are all from the normal peripheral zone, while all of the diseased states lie to the left of zero. When data is acquired from a new patient, in terms of T₁ and T₂ in units of ms and ADC in units of mm²/s, this point can be projected into the one-dimensional space using the values of the first principal component found from this experiment, and given a score with the simple equation, score =-0.000939*ADC -0.001342*T₁ -0.006772*T₂+4.79. The last term is a shift to center the data around zero.PCA is a powerful tool in the analysis of a data set that describes the directions in which maximal variance is observed. PCA analysis of the prostate data results in a simple equation that can be applied to a new data point consisting of the T₁, T_2, and ADC measurements, projecting it into the same one-dimensional space shown in Figure 1. This will allow for a fast additional test that can potentially rule out disease or flag the individual for further testing. While a strict delineation between the diseased states themselves is not evident, that is, we do not see a distinct clustering of low grade from high and intermediate grade prostate cancers, this is something that we are hoping for in future studies. This could be achieved either from using other clustering techniques, such as kernel PCA, or by expanding the dimensionality of the data to include additional quantitative parameters measured from MRF such as perfusion. In addition, more accurate data provided by the targeted biopsies on more patients, will improve the mathematical results.We have presented a simple method to cluster MRF data from the prostate into normal and diseased groups using principal component analysis. The single dimensional score shown here can be used to directly differentiate diseased from healthy tissues.