Optimization of noise reduction filtering for prostate cancer localization

Prostate cancer is one of the most common types of cancer. For many years, Magnetic Resonance (MR) Imaging has been frequently used by doctors in order to diagnose prostate cancer and follow-up stages. Artificial Intelligence (AI) based structures that have started to be included in the studies thanks to the developing technology in the recent period, aim to reduce the time period of diagnosis and follow-up stages of patients by creating an alternative diagnostic methods and to eliminate misdiagnosis by providing decision support structure for doctors. In these structures, Multi-Spectral (MS) MR images are frequently used to diagnose prostate cancer. However, noise in raw images complicates image processing. For this reason, MR images should be filtered as a pre-process before introduced into the diagnostic aid systems. In studies, filtering is often included in the structures as a pre-process, if not; the effect of the noise on the final result is tried to be reduced through an alternative model choice that can cope with noise. It is seen that, in studies which involve filtering as a pre-process, filter preferences are generally based on the experience in the field. Furthermore, selected filtering methods have simple forms in general. In this study, a novel method is developed that can select filter and filtering parameters to maximize cancer localization performance depending on the output of the selected decision mechanism. Moreover, it uses more complex filtering methods as compared to common use. It is aimed to eliminate the filter preference based on experience and to provide optimum filter preferences with measurable metrics. In this study, several spectrum are used, which are T2 Map generated from Spin Eco (SE);Apparent Diffusion Coefficient (ADC) from Diffusion Weighted Imaging (DWI); Initial Area Under the Curve (IAUC) 30, IAUC 60, IAUC 90 from Dynamic Contrast Enhanced Imaging (DCE) and kel, kep from the use of DCE on Brix model. 3 filters (Wiener (W)filter, Total Variant (TV) filter and Wavelet Thresholding (WTh)) and one parameter from each filtering are selected to be used on those spectra. Linear Discriminant Analysis (LDA) is used as a decision making model for cancer localization and selected filters along with their parameters are analyzed based on accuracy on cancer localization performance. It is observed that the selected filtering parameters have improved the cancer localization performance and a generic structure is established to be used for MS MR filtering optimization.