In contrast to the traditional systematic prostate biopsy that obtains multiple cores in a predefined pattern, a targeted biopsy involves only suspicious targets identified by a radiologist on multiparametric MRI (mpMRI) scans. These targets are biopsied either by an in-bore MR biopsy or by using MR/ultrasound fusion biopsy, said presenter Karthik Sarma, an MD-PhD student at the University of California, Los Angeles (UCLA).
"However, the identification of regions of suspicion on MRI is very challenging, as it requires the clinician to integrate information across multiple types of scans (generally T2, diffusion-weighted, and dynamic contrast-enhanced images)," Sarma told AuntMinnie.com. "My project aims to assist clinicians in this process by creating a deep-learning model that automatically analyzes mpMRI and produces a voxel-by-voxel 'cancer probability map' that can then be used when delineating regions of interest for biopsy."
The UCLA researchers used a dataset comprised of mpMRI scans from 82 patients with annotations created by expert prostate radiologists and histopathology-confirmed Gleason labels. With this dataset, they trained a four-layer convolutional neural network to classify mpMRI voxels as containing clinically significant prostate cancer -- defined as a Gleason score of 7.
"Using 10-fold cross-validation, our preliminary classifier achieved [an area under the receiver operator characteristics curve] of 0.78, which we have found very encouraging," Sarma said.
The team is now working to further improve the performance of the classifier by adding more patients to its dataset, and by making use of labels generated from whole-mount prostatectomy specimens, Sarma said.
Take in this poster at RSNA 2016 to learn more.
