Karen Elizabeth Burtt BS, Abstract Co-Author: Nothing to Disclose

Shijun Wang, Abstract Co-Author: Nothing to Disclose

Baris Turkbey MD, Abstract Co-Author: Nothing to Disclose

Evrim Bengi Turkbey MD, Abstract Co-Author: Nothing to Disclose

Nicholas Petrick PhD, Abstract Co-Author: Nothing to Disclose

Peter L. Choyke MD, Abstract Co-Author: Researcher, Koninklijke Philips NV Researcher, General Electric Company Researcher, Siemens AG Researcher, iCAD, Inc Researcher, Aspyrian Therapeutics, Inc Researcher, ImaginAb, Inc Researcher, Aura

Peter Pinto, Abstract Co-Author: Nothing to Disclose

Bradford J. Wood MD, Abstract Co-Author: Researcher, Koninklijke Philips NV Researcher, Celsion Corporation Researcher, BTG International Ltd Researcher, , W. L. Gore & Associates, Inc Researcher, Delcath Systems, Inc Pending research funded, Perfint Healthcare Pvt Ltd Patent agreement, VitalDyne, Inc Intellectual property, Koninklijke Philips NV Intellectual property, BTG International Ltd

Ronald M. Summers MD, PhD, Presenter: Royalties, iCAD, Inc Research funded, iCAD, Inc Stockholder, Johnson & Johnson Grant, Viatronix, Inc

We demonstrate an automated, supervised prostate computer-aided diagnosis (CADx) system utilizing a novel algorithm based on dynamic conditional random fields (DCRF) and evaluate the performance of this system with and without the use of a random forest (RF).

Multi-parametric 3T prostate MRI scans were performed on 60 patients using an endorectal coil. Pathology was established using MRI-TRUS fusion biopsies. Cancer and central gland segmentations were established by a trained radiologist for 40 training and 20 test cases on T2 weighted images (T2WI). ADC images and Ktrans maps (from DCE images) were registered to the T2 images using coordinate information. Features included intensities of T2WI, ADC, and Ktrans images, location information, and entropy of 3D sub-volume extracted around each node. A dynamic conditional random field with one layer of observations and two hidden layers was utilized to label and segment prostate lesions. The hidden variables defined whether a voxel was located in the central gland or in the peripheral zone, and whether it represented cancer. The DCRF classifier was trained using pseudo-negative log likelihood. A cascading classifier system was tested with an RF feeding into the DCRF. Prediction maps were generated by applying the classifier to test images. Statistical analysis of receiver operating characteristic (ROC) curves were conducted using the Mann-Whitney U Test.  

The proposed DCRF method without RF yielded an AUC of 0.8432 (95% C.I. [0.8422, 0.8441]) for classifying prostate lesions, compared to an AUC of 0.8889 (95% C.I. [0.8881, 0.8897]) using the DCRF with RF. DCRF with RF was found to have a statistically superior AUC compared to DCRF alone (p<0.01). Prediction maps and ROC curves are shown in Figure 1.

Random forest increases the performance of the proposed DCRF method in classifying cancer on multi-parametric prostate MRI. Clinically useful prediction maps may be generated using the proposed methods.

An improved CADx system may reduce reading time, increase the performance of less expert readers, and decrease inter-observer variability in interpreting prostate MR images.

Burtt, K, Wang, S, Turkbey, B, Turkbey, E, Petrick, N, Choyke, P, Pinto, P, Wood, B, Summers, R, Use of Random Forest in a Novel Dynamic Random Conditional Field-based Computer-aided Diagnosis System for Prostate Cancer Segmentation and Labeling on Multi-Parametric MR Images.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14019239.html