Premium
WE‐C‐116‐11: Prediction of Prostate Gleason Score Using Neural Network and Multi‐Parametric MRI
Author(s) -
Chen S,
D' Souza W,
Gullapalli R,
Mistry N
Publication year - 2013
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1118/1.4815573
Subject(s) - effective diffusion coefficient , receiver operating characteristic , standard deviation , percentile , nuclear medicine , medicine , prostate , prostatectomy , diffusion mri , magnetic resonance imaging , mathematics , radiology , statistics , cancer
Purpose: To predict prostate Gleason score (Gleason score >6) using neural network and multi‐parametric MRI. Methods: This study included 17 radical prostatectomy patients who underwent in‐vivo multi‐parametric MRI: Diffusion weighted images (DWI), T2‐Mapping, dynamic‐contrast enhanced (DCE) images, and MR spectroscopy (MRS) prior to surgery. The prostate gland was transversely sectioned and each tissue slice was radially divided into 8 sections. Two hundred thirty one of these sections were examined by a pathologist and a Gleason score was graded individually. For each section, distributions of corresponding parameters extracted from MRI images were used in our analysis. These parameters included the apparent diffusion coefficient (ADC) from DWI, T2‐map, forward volume transfer constant (Ktrans) from DCE, and MRS. The mean and standard deviation of the ADC, T2‐map, Ktrans, and their 95th, 90th, 10th, and 5th percentile values were derived from their distributions. A feed‐forward neural network was constructed using a growth algorithm. In this method, among the above variables, only the variables which increased the model performance were selected as input features of network. The network model was validated using a five‐fold cross‐validation Method: data were randomly divided into 5 groups of approximately equal size; four groups of data were used to train network and the remaining one was used to test the network, in turn. The model performance was evaluated with area under the Receiver Operating Characteristics (ROC) curve. Results: MRS, mean value of 90th percentage of Ktrans, mean and standard deviation of 10th percentage of ADC, and mean of 10th percentage of T2‐map were five predictors selected by the neural network model. The area under the ROC curve was 0.77 for cross‐validation data (sensitivity=72%, specificity=71%). Conclusion: Prostate Gleason score can be accurately predicted using multi‐parametric MRI and a neural network model. This provides a non‐invasive way to identify the prostate cancer aggressiveness.