Computer aided diagnosis of prostate MR
From UMCN Radiology Research
General project information
- Title: Computer assisted diagnoses of prostate cancer combining high resolution, dynamic contrast enhanced and spectroscopic MR
- Acronym : KUN 2004-3141
- Time frame: 2006 - 2009
- Funded by NWO as part of The Nijmegen Health Care Computing Initiative, hefboom program, Pr. Nr. KUN 2004-3141
Introduction
Currently available prostate cancer (pCa) staging and localization protocols require strong improvement. Up to 60\% of men have advanced disease after surgical resection, which may even worsen if the inflow of downstaged men due to positive PSA tests rises (14% of dutch men over 40 were tested in 2001). Localization accuracy is in the order of 60% which will seriously hamper effective application of new radio-therapy techniques such as IMRT. Recent magnetic resonance (MR) research has shown significant improvements (90\% staging and 90\% localization accuracy). These MR techniques however, are not clinically available on current MR systems, require expert knowledge and are still maturing. Purpose of this study is to research, implement and assess staging and localization techniques with MR such that it can run on any good quality MR system, requiring less expertise, and with optimized performance. Methods
MR provides three pCa relevant techniques: high resolution T2 weighted imaging (HRMR), dynamic contrast enhanced MR (DCEMR) and proton MR Spectroscopy (HMRS). We have been applying and improving state-of-the-art HRMR, DCEMR and HMRS to hundreds of pCa patients over 10 years. In retrospective studies using our large institutional database of whole mount section pathology several MR features were demonstrated to correlate well with pathology. In recent subsequent prospective studies we were able to demonstrate the clinical feasibility of MR in staging and localization pCa. We have identified 4 topics requiring further research before MR pCa diagnosis can be effectively implemented in clinical practice. They are: 1. Substantial interpretation variability occurs due to the large number of and the natural heterogeneity of the various anatomical and functional features, 2. Feature fluctuations occur due to failing calibration techniques. 3. Feature inhomogeneity artifacts due to coil profile, 4. Patient movement artifacts.
A dual approach is proposed to ensure the clinical usability of the project. The physics research solutions are concurrently evaluated in prospective studies on pCa patients regularly scheduled for MR. We will first start to research (and evaluate) opportunities to reduce interpretation variability. Automated multi-feature pattern recognition techniques will be considered that are trained on already existing databases to objectively extract information regarding detection/location, density and likelihood of pCa. Feature extraction techniques will be evaluated that reproducibly characterize a lesion/foci by its most aggressive part. Secondly, feature estimation methods will be improved by including more prior knowledge, and developing robust clinically feasible calibration techniques. Thirdly, coil inhomogeneity correction techniques will implemented and evaluated. Finally, image registration techniques will be included and evaluated for correcting artifacts due to patient movement.
Results
- Prototype CAD system for multi-modal prostate MR
- Screenshot of CAD application visualizing multi-modal MR information. Top left is a slice from a T2 weigted transversal volume. Top middle is the same t2-w image both with the pharmacokinetic Ve (extravascular extracellular volume) parameter in overlay. Top right same as middle, but then for Ktrans (permeability). Bottom left proton MR spectrum at the indicated location. Bottom middle, dynamic contrast enhancment curve at the specified location. Bottom right t1-w transversal slice with HMRS ratio overlay.
- Diagnostic performance evaluation of pharmacokinetic DCE-MRI based prostate CAD
- A novel automated computerized scheme has been developed for determining a likelihood measure of malignancy for suspicious regions in the prostate on dynamic contrast-enhanced sequences. Our database consisted of 120 patients with histologically proved adenocarcinoma of the prostate. Both carcinoma and normal tissue were annotated by a radiologist and researcher using histopathology. The annotated regions were used to select a region of interests (ROI) in de dynamic contrast-enhanced sequences. A two-compartment pharmacokinetic feature set was extracted from the ROIs to train a linear discriminant classifier. We used the output of the classifier as a measure of likelihood of malignancy. General performance of the scheme was evaluated using the Az value (area under the curve). Results showed a discriminating performance of 0.83. The automated computerized scheme is therefore an asset for the radiologist to detect and diagnose prostate cancer.