Assessing the Impact of a Pre-Processing Pipeline on a Tumor Localization Model in Prostate MRI Within a Large Multi-Institutional Dataset (NRG-GU005)

Stephanie Alley, Marion Tonneau, Damien Olivié, Clare M. Tempany-Afdhal, Peter L. Choyke, Baris I. Turkbey, Uulke A. van der Heide, Rodney J. Ellis, Thomas P. Boike, J. Daniel Pennington, Arthur Frazier, Colleen A. F. Lawton, Nelson Leong, Alina M. Mihai, Scott C. Morgan, Abhishek A. Solanki, Jeff M. Michalski, Felix Y. Feng, Howard Sandler, Cynthia Menard, Samuel Kadoury.

Assessing the Impact of a Pre-Processing Pipeline on a Tumor Localization Model in Prostate MRI Within a Large Multi-Institutional Dataset (NRG-GU005)

Proceedings Volume 13410, Medical Imaging 2025: Clinical and Biomedical Imaging; 1341004 (2025).

Abstract

Multi-parametric magnetic resonance imaging (mpMRI) is increasingly recognized as a valuable tool for characterizing prostate cancer, integrating T2-weighted (T2w), diffusion-weighted (DWI), and dynamic contrast-enhanced (DCE) imaging. Despite its high sensitivity in localizing tumors, the specificity of mpMRI was shown to be hindered by benign conditions that mimic cancerous tissue features. The aim of this study is to investigate the effect of a pre-processing pipeline integrating state-of-the-art registration, bias field correction and normalization tools. We validated this pre-processing pipeline on a large multi-site dataset of 468 patients from 109 institutions. After performing all pre-processing, tumor localization was determined using a model-based tumor localization approach that takes both multi-parametric MRI and prior clinical knowledge features as input. Results show deformable image registration yielded a significant improvement in tumor localization accuracy, both for the diameter analysis as well as the area under the curve comparison for the subset of patients with ground truth tumor delineations.