Multicenter validation of prostate tumor localization using multiparametric MRI and prior knowledge.

Area of Science:

• Medical Imaging
• Radiotherapy
• Oncology

Background:

• Accurate tumor localization is vital for radiotherapy, enabling precise dose delivery and minimizing side effects.
• Multiparametric MRI (mp-MRI) aids prostate cancer detection and localization by visualizing tissue characteristics.
• Distinguishing cancer, especially in the prostate's transition zone, remains a challenge with mp-MRI alone.

Purpose of the Study:

• To enhance the performance of mp-MRI-based prostate tumor localization models.
• To integrate prior knowledge from population-based tumor probability atlases and patient-specific biopsy results.
• To improve upon physician-based manual tumor delineation.

Main Methods:

• A cohort of 40 patients from two centers underwent mp-MRI (T2-weighted, diffusion-weighted, dynamic contrast-enhanced).
• 31 features were extracted per voxel, including 29 from mp-MRI, 1 from a tumor probability atlas, and 1 from biopsy data.
• Model performance was validated in single-center and cross-center settings, comparing automated delineations to manual ones.

Main Results:

• Integrating prior knowledge features significantly improved the area under the ROC curve (AUC) from 0.690 to 0.775 in the single-center setting.
• The combined model achieved performance comparable to the average of six manual delineations by expert teams, with an error rate of 0.22.
• Cross-center validation also showed robust performance (mean AUC of 0.777), indicating generalizability.

Conclusions:

• Incorporating prior knowledge features from tumor atlases and biopsies substantially boosts the accuracy of mp-MRI-based prostate tumor localization.
• The developed model demonstrates effectiveness in both single-center and cross-center validation, outperforming mp-MRI alone.
• This approach offers a promising tool for enhancing radiotherapy planning and treatment.