Encrypted login | home

Program Information

Prostate Cancer Foci Detection and Aggressiveness Identification Using Multi-Parametric MRI/MRS and Supervised Learning


G Kirlik

G Kirlik*, W D'Souza , M Naslund , J Wong , R Gullapalli , H Zhang , University of Maryland School of Medicine, Baltimore, MD

Presentations

TH-CD-207-7 (Thursday, July 16, 2015) 10:00 AM - 12:00 PM Room: 207


Purpose: To develop a non-invasive and quantitative information platform to accurately identify the prostate cancer foci and tumor aggressiveness grade.

Methods: 11 patients who received magnetic resonance imaging (MRI) and magnetic resonance spectroscopy imaging (MRSI) prior to the surgery were studied. Multi-parametric imaging included extracting T2-map, Apparent Diffusion Coefficient (ADC) using diffusion weighted MRI, Ktrans using Dynamic Contrast Enhanced MRI, and 3D-MR Spectroscopy. Each image was composed of approximately 10 slices, which were divided into octants and individually assessed for the presence/absence of tumor by a radiologist. Following the radical prostatectomy, digital images of both the slice specimens and the histopathology slides were obtained. A pathologist reviewed all 223 octants and marked cancerous regions on each and graded them with Gleason score, which served as the ground truth to validate our prediction. Both binary prediction (indolent/aggressive cancer, separated by Gleason = 6) and multi-class prediction (no-cancer, non-aggressive cancer and aggressive cancer) were performed. Adaptive boosting with random under-sampling method was used. Area under the receiver operating characteristic (ROC) curve (AUC) and 95% confidence intervals were collected after repetitions of 10-fold cross-validation.

Results: In binary prediction, average AUC value of was 0.77 [0.73, 0.79] with average sensitivity and specificity 79% [74%, 85%] and 74% [70%, 77%], respectively. For the multi-class classification, the average prediction accuracy for the no-cancer class was 97% [91%, 100%], for non-aggressive cancer class was 80% [76%, 83%] and for aggressive cancer class was 65% [59%, 72%]. The overall accuracy was 86% [83%, 88%].

Conclusion: We provided a sophisticated while user-friendly platform using multi-parametric MRI combined with supervised learning to be able to accurately detect cancer foci and its aggressiveness. In addition, our method was non-invasive and allowed for non-subjective disease characterization, which provided physician information to make personalized treatment decision.


Contact Email: