ABSTRACT
PURPOSE: Tumor localization provides crucial information for radiotherapy dose differentiation treatments, such as focal dose escalation and dose painting by numbers, which aim at achieving tumor control with minimal side effects. Multiparametric (mp-)MRI is increasingly used for tumor detection and localization in prostate because of its ability to visualize tissue structure and to reveal tumor characteristics. However, it can be challenging to distinguish cancer, particularly in the transition zone. In this study, we enhance the performance of a mp-MRI-based tumor localization model by incorporating prior knowledge from two sources: a population-based tumor probability atlas and patient-specific biopsy examination results. This information typically would be considered by a physician when carrying out a manual tumor delineation. MATERIALS AND METHODS: Our study involves 40 patients from two centers: 23 patients from the University Hospital Leuven (Leuven), Leuven, Belgium and 17 patients from the Netherlands Cancer Institute (NKI), Amsterdam, the Netherlands. All patients received a mp-MRI exam consisting of a T2-weighted, diffusion-weighted, and dynamic contrast-enhanced MRI before prostatectomy. Thirty-one features were extracted for each voxel in the prostate. Among these, 29 were from the multiparametric-MRI, one was from the population-based tumor probability atlas and one from the biopsy map. T2-weighted images of each patient were registered to whole-mount section pathology slices to obtain the ground truth. The study was validated in two settings: single-center (training and test sets were from the same cohort); and cross-center (training and test sets were from different cohorts). In addition, automatic delineations created by our model were compared with manual tumor delineations done by six different teams on a subset of Leuven cohort including 15 patients. RESULTS: In the single-center setting, mp-MRI-based features yielded area under the ROC curves (AUC) of 0.690 on a pooled set of patients from both cohorts. Including prevalence into mp-MRI-based features increased the AUC to 0.751 and including all features achieved the best performance with AUC of 0.775. Using all features always showed better results when varying the size of the training set. In addition, its performance is comparable with the average performance of six teams delineating the tumors manually. The error rate using all features was 0.22. The two prior knowledge features ranked among the top four most important features out of the 31 features. In the cross-center setting, combining all features also yielded the best performance in terms of the mean AUC of 0.777 on the pooled set of patients from both cohorts. In addition, the difference in performance between the single-center setting and cross-center setting was not significant. CONCLUSIONS: The results showed significant improvements when including prior knowledge features in addition to mp-MRI-based features in both single- and cross-center settings.
