Topology of DCE MR Perfusion Pattern Within Cervical Cancers During Radiation Therapy

Abstract

Purpose/Objective(s): To develop a non-invasive and quantitative information platform to accurately identify the prostate cancer foci and tumor aggressiveness grade. Materials/Methods: Eleven patients (PSA score ranged from 0.5 to 29.0 with an average 9.4) who had biopsy-proven PCa and elected to have radical prostatectomy received magnetic resonance imaging (MRI) and magnetic resonance spectroscopy imaging (MRSI) prior to the surgery. Multi-parametric imaging included extracting a) T2-map, b) Apparent Diffusion Coefficient (ADC) using diffusion weighted MRI, c) K using Dynamic Contrast Enhanced MRI, and d) 3D-MR Spectroscopy using 3D PRESS covering the entire prostate. Each imagewas composed of approximately 10 slices, which were divided into octants and individually assessed for the presence/absence of tumor by a radiologist (resulting in w223 octants that were usable). Following the radical prostatectomy, digital images of both the slice specimens and the histopathology slides were obtained. A pathologist reviewed all 223 slides and marked cancerous regions on each slide and graded them with Gleason score. The pathologist’s grade for the samples served as the ground truth to validate our prediction. Imaging parameters were combined with supervised learning techniques to provide the informatics for diagnosis (binary prediction: cancer/no cancer or indolent/ aggressive cancer, separated by Gleason Z 6) and for treatment decisions (categorical prediction: actual aggressiveness grade). Prediction platform was constructed based on a hybrid of three supervised learning approaches, namely logistic regression, neural network and Naive Bayes. Results: Binary diagnosis prediction combinedwith the radiologist read as an input to the platform, the hybrid method achieved an area under the receiver operating characteristic (AUC) of 0.8. This was higher than any single supervised learning approach 0.67, 0.77 and 0.66. For categorical classification of tumor aggressive grade the hybrid method achieved an AUC of 0.7 as opposed to the individual approachAUCsof 0.65, 0.66 and0.63. Ifweuse only the imaging parameters as input, we were able to double the sensitivity of the radiologist’s diagnosis while maintaining the same specificity. Conclusions: A single imaging modality or conventional MRI alone lacked the sensitivity and specificity to identify prostate cancer foci. We provided a sophisticated while user-friendly platform using multi-parametric MRI combined with hybrid 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. Author Disclosure: H.H. Zhang: E. Research Grant; Varian Medical Systems. N. Mistry: None. D. Anderson: None. B. Golden: None. M. Naslund: None. J. Wong: None. R. Gullapalli: None. W. D’Souza: E. Research Grant; Varian Medical Systems.