MRCAD for Daily Clinical Analysis of Prostate MR
The Diagnostic Image Analysis Group (DIAG) from the Department of Radiology at Radboud University Nijmegen Medical Centre in The Netherlands researches computer-aided diagnosis of breast, lung and prostate cancer. The research in prostate cancer is a collaborative effort with the clinical prostate group in the department and is focused on high resolution anatomical, functional (diffusion and dynamic contrast enhanced) and metabolic (spectroscopy) MR. During various prostate and breast MR projects an image analysis workstation, MRCAD, has been gradually developed based on VTK, ITK and IncrTCL. This workstation provides dedicated image analysis tools and display facilities for various state-of-the-art MR sequences and hardware as well as a platform for testing new computerized image analysis techniques. Currently, MRCAD is in daily clinical use at our clinic and at a few other sites. MRCAD is being used to reporte prostate MR and in various research projects and publications. This article will address some of the design features and explain how VTK, ITK and IncrTCL have been used to achieve a robust, versatile, clinical research platform.
Screenshots
To help understand MRCAD, here are two example cases
The above figure shows an MR detection study of a patient with elevated PSA, where prior ultrasound biopsy could not confirm the presence of cancer. The detection protocol includes: t2-w, ADC, and DCE-MRI sequences. The radiologist marks a finding, which is suspicious for cancer, with a transparent red sphere (vtkUnstructuredGrid) and is presented with CADx output to assist in differentiating between benign and malignant. A malignancy score of .33 (green and red are scores of benign and malignant training findings) indicates that the region is unlikely to contain cancer. This patient is currently in an active surveillance program.
A 59-year-old patient with a PSA of 12 ng/ml and a confirmed biopsy Gleason 8 tumor in the right peripheral zone. The MRCAD hanging protocol shows T2-w transversal (right view) and in color overlays: Pharmacokinetic DCE-MRI; and ADC-map (left upper row); and choline metabolite concentration; coronal T2-w image (left bottom row); The separate window shows a time-concentration curve and an MR Spectroscopy spectrum at the cursor location. The reporting radiologist rated all 4 modalities 5 on a 5 point scale: where 1 is no tumor, and 5 is definitely tumor. These MR findings were confirmed retrospectively with a stage T3a determined after prostatectomy.
Overview of the software
The workstation runs an OpenSuse Linux OS. A DCMTK server daemon receives DICOM images automatically forwarded by the PACS system. A polling mechanism triggers automatic pre-processing of MR studies (e.g., fitting DCE-MRI curves). The studies are then ready for viewing and analysis. MRCAD features are: compressed, multiframe DICOM Reading/Writing; multi-modal color overlay; DCEMR pharmacokinetic modeling and registration; manual and semi-automatic segmentation; region of interest statistics; CADx, CADe; MRS (using LCModel), and Structured Reporting.
The MRCAD software has a layered structure as shown below. The main programming language is Tcl/Tk with the object oriented extension IncrTcl. Processing and display are performed in VTK pipelines that are setup using the Tcl wrapped interface. Local VTK classes provide additional high computational functionality at the C++ level or as means to include wrappable ITK functionality.
Annotation and Structured Reporting
Clinical findings and ground truth for CAD are annotated and stored in XML. Example tags are: tumor location, tumor grade, scorings, and access time. For structured reporting, findings can be collected and added to the XML database. A PDF file can be generated from the XML file (using LaTeX) and, after inspection, sent to an Eletronic Patient Database using SOAP requests. The XML data is additionally used for training CAD systems, and performing clinical observer studies. An XML example of an observer scoring is:
PatientID="12345" StudyDate="20100217"
InstitutionName="UMC_St_Radboud"
StudyDescription="ABDOMEN-ONDERBUIK">
<marklist username="pieter" mode="detect" repeat="1" studysetname="Default">
<mark id="mark1">
<Time>2010.02.24-11:51:05</Time>
<Position>-25.7904 9.30513 -53.7623</Position>
<AlignedModality>BGT2Tra1</AlignedModalityBG>
<Type>Tumor</Type>
<T2>5</T2><DCE>4</DCE><DIFF>4</DIFF>
<Link id="Link1">
<StudyDate>20091220</StudyDate>
</Link>
</mark>
<markpatient id="patientlevel">
<Stage>T2c</Stage>
<!--prostate size -->
<AP>46</AP><RL>57</RL><CC>46</CC>
</markpatient>
</marklist>
</markdatasetlists/>
The node markdatasetlists is the root node for all annotations of the MR study. In this example (user 'pieter' in the marklist node) a tumor is detected and is rated 5-4-4 on 3 MR modalities. Furthermore, the finding is linked to another finding in a prior MR study (Link1). The patient diagnosis is in the markpatient node: a T2c tumor. After reporting the patient case, MRCAD can use the findings stored in the XML database to collect screenshots and format the information into a PDF file using dedicated stylesheets. This PDF report is then sent to the Electronic Patient Database using SOAP:
boundary=''" -headers [list Cookie $sessionid] -query [$soaprequest asXML]]
set body [http::data $token]
http::cleanup $token
Coordinate system
MRCAD uses three coordinate systems to handle the variety of image volume and voxel dimensions and orientations in an MR scan: ijk voxel location in an MR series, xyz patient position, uvw viewport coordinate. A vtkDCMTKImageReader directly reads DICOM images in single slice, multiframe, and/or compressed format and produces a vtkImageData and a vtkTransform (ijk2xyz). The latter transform maps voxel location ijk to an xyz patient position. The viewer uses an additional uvw2xyz transform to handle: zooming, panning and slice selection. This allows for accurate (sub-voxel) overlay of any image (e.g transversal 0.5x0.5x3mm) onto any other orientated image (oblique (15deg) sagital (1x1x4mm)). Arbitrarily shaped 3D volume annotations are created and stored in xyz coordinates using the vtkUnstructuredGrid format. These annotations can be added and overlayed to any view and are also used to compute region of interest statistics in the various multi-modal images.
Data and widget factory
The core of the MRCAD application is designed as a factory method pattern. The approach ensures fast initializing of the application, by constructing data objects on demand. As a result, VTK pipelines are only created when needed. An IncrTcl coded part of the factory:
...
public {
# static method that returns the singleton proc Instance {}
# check whether the data object ''mapname'' is registered
method IsAvailable {mapname}
# Register the data object named ''mapname'' to the factory
method AddData {mapname data}
# Returns the data object from the factory
method GetData {mapname}
...
}
}
# Implementation
body MRCADDataFactory::GetData {mapname} {
set data $_rgContainer($mapname)
set object [find object *$data]
if { $object == ""} {
# create the object only once
set object [[lindex $data 0] \#auto [lrange $data 1 end]]
set _rgContainer($mapname) $object
}
return [namespace which $object]
}
The factory is implemented as a singleton and can be queried using the static member function Instance. The factory is filled with possible data during MRCAD initialization, for example a diffusion-w series might be present, then:
[MRCADDataFactory::Instance] AddData "Diff1"
[list DiskImageData $files]
adds an imagedata object with mapname "Diff1", which is constructed by a specialized class DiskImageData (an IncrTcl wrapper of the vtkDCMTKImageReader).
To view the diffusion-w image in the application:
"Diff1"} {
# Get the data object and create the VTK pipeline nescessary to produce the data
set obj [[MRCADDataFactory::Instance] GetData "Diff1"]
# Display the image in the viewer
vlSetImage BG [$obj GetVolume]
vlSetTransform BG [$obj GetTransform]
}
Data objects can be accompanied by data widgets to allow user interaction with the data object. In the image below, the spectroscopic metabolite image data object output is displayed as a transparent, color-coded overlay on top of a T2-w image. Multiple metabolite maps are available and the accompanying data widget allows selection of the metabolite and provides other interactions with the data object. Here, the 3th available metabolite, choline (Cho), is selected:
Research overview
MRCAD is used in a number of research projects. MRCAD was used to detect and annotate the Dominant Intraprostatic Lesion for radiotherapy IMRT planning [1]. A successful Computer Aided Diagnosis system has been researched and implemented to discriminate benign from malignant suspiscous regions using region of interest statistical features derived from DICOM images directly or processed parametric maps [2]. Pattern recognition as well as ROC analysis was performed using the statistical package R. Pharmacokinetic features computed from Dynamic Contrast Enhanced MRI provided the highest diagnostic accuracy. Subsequently in [3], more modalities/features were added to improve the discriminating performance. An ITK-based registration method was extended with a local incompressibility term to add MR series features that were misaligned by possible patient movement during acquisition. Currenty, [4] and [5] an initial system for fully automatic detection of prostate cancer is being developed using a locally developed vtkClass based on the itkHessianRecursiveGaussianImageFilter to detect lesions. MRCAD is also used in several observer studies that are performed by radiologists to determine the diagnostic value of a certain MR modality or protocol. In [6] MRCAD was used for scoring DCE-MRI derived parameters and establish the value for localizing prostate cancer. Recent use includes detecting MR suspicious lesions for MR or ultrasound guided biopsy and detecting recurrence after radiotherapy [7,8].
References
[1] Prostate cancer: precision of integrating functional MR imaging with radiation therapy treatment by using fiducial gold markers
(Huisman, et al. Radiology, 2005)
[2] Computerized analysis of prostate lesions in the peripheral zone using dynamic contrast enhanced MRI (Vos, et al. Medical Physics 2008)
[3] Computer assisted analysis of peripheral zone prostate lesions using T2-weighted and dynamic contrast enhanced T1-weighted MRI
(Vos, et al. PMB 2010)
[4] Automated calibration for computerized analysis of prostate lesions using pharmacokinetic magnetic resonance images
(Vos, et al. MICCAI 2009)
[5] Computer-aided Detection of Prostate Lesions at Diffusion-weighted MR Using a Dedicated Hessian Matrix–based Detection Scheme
(Vos et al. RSNA, 2009 Chicago)
[6] Prostate cancer localization with dynamic contrast-enhanced MR imaging and proton MR spectroscopic imaging
(Futterer, et al. Radiology 2006)
[7] Magnetic resonance imaging guided prostate biopsy in men with repeat negative biopsies and increased prostate specific antigen (Hambrock, et al. Journal of Urology 2010)
[8] Feasibility of 3T Dynamic Contrast-Enhanced Magnetic Resonance-Guided Biopsy in Localizing Local Recurrence of Prostate Cancer After External Beam Radiation Therapy (Yakar, et al. Investigative Radiology 2010)
Henkjan Huisman
is an Assistant Professor in the Diagnostic Image Analysis Group in the Department of Radiology at Radboud University Nijmegen Medical Centre. His research interests include computer aided diagnosis, image segmentation, prior knowledge integration in image analysis, and pharmacokinetic MRI.
Pieter Vos
is a PhD student at the Diagnostic Image Analysis Group in the Department of Radiology at Radboud University Nijmegen Medical Centre where he's working on Computer Aided Diagnosis of prostate cancer using multimodal MR.
This work is licensed under a Creative Commons Attribution-NoDerivs 3.0 Unported License. Additional information on this license can be found here.