Preoperative neural network using combined magnetic resonance imaging variables, prostate-specific antigen, and Gleason score to predict positive surgical margins

Objectives. To assess whether artificial neural network analysis (ANNA) predicts for positive surgical margins (PSMs) more effectively than logistic regression analysis (LRA) according to the combined use of the findings of pelvic coil magnetic resonance imaging (pMRI) and other preoperatively available tumor variables in patients with clinically organ-confined prostate cancer after radical prostatectomy. Methods. A total of 205 patients with clinically localized prostate cancer, who underwent retropubic radical prostatectomy were evaluated. The predictive variables included clinical TNM stage, prostate-specific antigen (PSA) level, PSA density, biopsy Gleason score, percentage of cancer in biopsy specimens, and pMRI findings. The predicted outcome was PSMs. The patient data were randomly split into four cross-validation sets and used to develop and validate the ANNA and LRA models. For comparison, the area under the receiver operating characteristic curve was used. Results. The overall PSM rate was 22% (n = 45). Using all input parameters, the accuracy of the ANNA and LRA was 84% and 75% for the prediction of PSMs, respectively. The area under the receiver operating characteristic curve of the ANNA (0.872 +/- 0.014) was significantly greater statistically (P <0.001) than that for LRA (0.791 +/- 0.006). The simplified ANNA. models that used the pMRI findings in addition to PSA and Gleason score were as accurate as the model that used all the variables (P = 0.89). A high percentage of cancer in the biopsy specimens, pMRI findings, and high PSA density were equally the most influential predictors (relative weight 1.881, 1.964, and 1.493, respectively). Conclusions. All the ANNA models in this study were superior to LRA in the prediction of PSMs. The ANNA using pMRI findings, PSA level, and Gleason score as input variables performed as well as the ANNA using all the input parameters. Additional studies Seem warranted.