Diffusion MRI preprocessing affects ADC estimation and automatic PI-RADS v2.1 classification in bi-parametric prostate MRI

Jul 16, 2026

Diffusion-weighted imaging (DWI) is integral to bi-parametric and multiparametric prostate MRI, yet preprocessing artifacts can undermine quantitative accuracy and clinical interpretation. This study evaluates how denoising, Gibbs-ringing correction, and susceptibility distortion correction influence apparent diffusion coefficient (ADC) estimation and automatic PI-RADS v2.1 classification. Using 268 fastMRI-derived cases, the work contrasts ADC fitting methods and trains a DenseNet-based 3-class deep model. Results show preprocessing changes ADC values and improves high-risk classification, supporting optimized prostate MRI pipelines.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages19

Reading Time48 min

Introduction

Methods

Results

Conclusion

Diffusion MRI preprocessing affects ADC estimation and automatic PI-RADS v2.1 classification in bi-parametric prostate MRI Christos Kanakis 1 ,3 , Mathias Perslev3 , Tim Schakel2 , Silvia Ingala3 ,4 ,5 , Akshay Pai3 , Dennis Klomp 1 , Chantal M.W. Tax 1 1 Center for Image Sciences, University Medical Center Utrecht, Utrecht, The Netherlands 2 Department of Radiotherapy, University Medical Center Utrecht, Utrecht, The Netherlands 3 Cerebriu, Copenhagen, Denmark 4 Department of Diagnostic Radiology, Copenhagen University Hospital Rigshospitalet, Copenhagen, Denmark 5 Department of Diagnostic Radiology, Copenhagen University Hospital Herlev and Gentofte, Herlev, Denmark arXiv :2607 . 11385v1 [ ee ss .IV] 13 Jul 2026 Abstract— Introduction: Diffusion-weighted imaging (DWI) is acquired as part of bi-parametric and multiparametric prostate MRI, but suffers from multiple artifacts that degrade downstream quantitative and diagnostic performance. WhileDWI preprocessing is more standard in brain imaging, its adoption in prostate imaging remains limited and lacks standardized pipelines. This study investigated the effect of different DWI preprocessing strategies on apparent diffusion coefficient (ADC) estimation and automatic Prostate Imaging Reporting and Data System (PI-RADS) classification performance. Methods: 268 cases were derived from the fastMRI prostate cohort by sequentially applying preprocessing steps: denoising, Gibbs-ringing correction, and diffeomorphic registration for susceptibility distortion correction, creating differently preprocessed datasets. First, ADC maps were compared using both linear least squares (LLS) and iterativelyweighted LLS (IWLLS) estimation. Next, a 3-class deep learning classifier based on a DenseNet architecture was trained to predict PI-RADS scores from multi-channel MRI inputs. Results: ADC analysis revealed statistically significant differences across preprocessing pipelines, with LLS and IWLLS estimation producing numerically equivalent ADC maps. Linear relationships between ADC values were preserved across most datasets (Pearson Correlation Coefficient, PCC ∼0.99), while distortion correction realigned the DWI to the T2w anatomy and altered ADC values accordingly (PCC ∼0.90). Automatic classification showed the best area under the receiver operating characteristic curve (AUROC) and sensitivity for high-risk PIRADS classes in the fully processed dataset (Dataset 5). False negative (FN) analysis revealed that Dataset 5 produced the least overconfident incorrect predictions on high-risk classes, a desirable property for clinical triage. Conclusion: DWI preprocessing, particularly the inclusion of distortion correction, enhances both the quantitative quality of ADC maps and the predictive power of deep learning models for PI-RADS classification. These findings support the need for optimized DWI preprocessing pipelinesin prostate MRI and highlight their potential to improve diagnostic accuracy and support automated triaging of high-risk prostate cancer. Keywords— diffusion, preprocessing, ADC maps, classification, PI-RADS, deep learning I. INTRODUCTION Prostate cancer (PCa) is the most common malignancy affecting the male population. Magnetic Resonance Imaging (MRI) is currently the cornerstone of diagnostic and prognostic assessment of PCa, offering the possibility of risk stratification [1] and facilitating the selection of biopsy candidates [2] in certain populations. The Prostate Imaging Reporting and Data System (PI-RADS) v2.1 [3] guides clinical decisions by assigning lesion scores from 1 to 5, reflecting the likelihood of clinically significant PCa (csPCa) . PI-RADS scores of 1 or 2 are considered low risk, PI-RADS 3 is assigned to indeterminate cases and PI-RADS 4 or 5 are considered high risk and warrant further investigation. However, workflows can vary significantly among different institutions[4]: Multiparametric (mp)-MRI protocols include acquisition

FAQs

Which DWI preprocessing strategies are evaluated in the study?

The study sequentially applies denoising, Gibbs-ringing correction, and diffeomorphic registration for susceptibility distortion correction, producing multiple preprocessed datasets.

How do the preprocessing pipelines affect ADC estimation?

ADC maps differ across pipelines; LLS and IWLLS yield numerically equivalent ADC maps, and distortion correction realigns DWI to T2w anatomy while altering ADC values.

What preprocessing setup improves automatic PI-RADS v2.1 classification performance?

The fully processed dataset (Dataset 5) achieves the best AUROC and sensitivity for high-risk PI-RADS classes and shows the least overconfident incorrect predictions for false negatives.

Jul 16, 2026

Upload by Genevieve