Automated Quality Control of Prostate Cancer Mrsi Using Independent Component Analysis

Introduction. The current diagnosis and localization of suspected prostate cancers requires TRUS-guided sextant (or more) biopsy with limited accuracy. MRI and MR Spectroscopic Imaging (MRSI) of prostate cancer patients can provide complimentary information to these techniques on the detection, localization and grading of prostate cancers with minimal risk to the patient [1]. However the spectra in MRSI data sets can contain artifacts, such as poor signal-to-noise, contaminating lipid signals and poor shim, which take expert knowledge to identify. If the analysis of MRSI data of prostate tumour is to be automated for clinical use then quality control (QC) will be a necessary step in that process. A previous method[2] of QC for prostate MRSI data at 1.5T used parameters of an applied quantification algorithm for quality control, such as the residual of the fit and the metabolite ratio that was also used for identifying tumour. We present a feature extraction method, using independent component analysis (ICA), that separates 3T MRSI data into acceptable and unacceptable quality groups using the raw spectra. We hypothesise that a feature extraction using independent components (ICs) generated from acceptable quality spectra will generate high scores for these spectra. Feature extraction of poor quality spectra will produce low scores for the same ICs as they will characterize these data less well giving a separation of spectra on the basis of quality.