Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/model.json

Name

Prostate Lesion Detection Using Multisite Biparametric MRI Datasets

Link

https://dx.doi.org/10.1148/ryai.230521

Indexing

Keywords: Prostate Cancer Detection, Multisite, Unsupervised Domain Adaptation, Diffusion-weighted Imaging, b value, bpMRI, ADC, DWI, PI-RADS

Content: GU, MR, OI

RadLex: RID12698, RID38890, RID50313, RID50294, RID50307

Author(s)

Hao Li

Han Liu

Heinrich von Busch

Robert Grimm

Henkjan Huisman

Angela Tong

David Winkel

Tobias Penzkofer

Ivan Shabunin

Moon Hyung Choi

Qingsong Yang

Dieter Szolar

Steven Shea

Fergus Coakley

Mukesh Harisinghani

Ipek Oguz

Dorin Comaniciu

Ali Kamen

Bin Lou

Organization(s)

Digital Technology and Innovation, Siemens Healthineers

Diagnostic Imaging, Siemens Healthineers

Vanderbilt University

Radboud University Medical Center

New York University

Universitätsspital Basel

Charité – Universitätsmedizin Berlin

Patero Clinic

Eunpyeong St. Mary’s Hospital, Catholic University of Korea

Changhai Hospital of Shanghai

Diagnostikum Graz Süd-West

Department of Radiology, Loyola University Medical Center

Oregon Health and Science University School of Medicine

Massachusetts General Hospital

Version

1.0

Funding

Authors declared no funding for this work.

Ethical review

Study approved by local ethics committees at all participating institutions; written informed consent obtained or waived as appropriate.

Date

Published: 2024-08-21

References

[1] Li H, Liu H, von Busch H, Grimm R, Huisman H, Tong A, Winkel D, Penzkofer T, Shabunin I, Choi MH, Yang Q, Szolar D, Shea S, Coakley F, Harisinghani M, Oguz I, Comaniciu D, Kamen A, Lou B. "Deep Learning–based Unsupervised Domain Adaptation via a Unified Model for Prostate Lesion Detection Using Multisite Biparametric MRI Datasets". Radiology: Artificial Intelligence. 2024;6(5):e230521.. 2024-08-21. doi:10.1148/ryai.230521. PMID: 39166972. PMCID: PMC11449150.

Model

Architecture

Unified unsupervised domain adaptation framework combining image synthesis generators adapted from CUT (U-shaped networks with dynamic filter for meta-information driven multidomain mapping) and a 2D U-Net with residual blocks for lesion detection.

Clinical benefit

Improves prostate cancer lesion detection performance and robustness across heterogeneous, multisite bpMRI acquisitions, particularly when DWI b-values deviate from PI-RADS recommendations.

Clinical workflow phase

Clinical decision support systems; workflow optimization for radiologist reading of prostate MRI.

Degree of automation

Automated image-to-image translation to reference domain and automated lesion heatmap generation; supports radiologist decision-making without additional labeling of target data.

Indications for use

Detection of prostate cancer lesions on biparametric MRI (T2-weighted + DWI) across multiple sites with varying DWI b-values in adult male patients, within radiology diagnostic environments.

Input

T2-weighted MR images, ADC maps, computed DWI at fixed b=2000 s/mm^2 (from available low/high b-value DWI), prostate mask, and meta-information (low/high b-values).

Instructions

Apply the unified generator to translate target-domain ADC and DWI B-2000 images into the reference-domain style using meta-information about b-values; then feed T2-weighted, generated ADC, generated DWI B-2000, and prostate mask into the pretrained 2D U-Net detection model. No retraining of the detection model is required.

Limitations

Domain indicator uses only b-values; performance may be suboptimal when high b-values closely resemble reference domain; imbalanced b-value distribution and limited samples in some groups affected performance; study focused on ADC and DWI high-b images (T2-weighted not adapted); vendor-provided ADC maps were excluded; generalization to other protocol factors (field strength, sequence, averages) not evaluated.

Output

CDEs: RDE1546, RDE1550

Description: PCa lesion heatmap and case-level prediction score; connected components derived for lesion candidates (true/false positives by overlap or distance to annotations).

Recommendation

Use UDA-generated images especially when target DWI b-values deviate from PI-RADS-recommended settings to improve detection robustness of pretrained models.

Regulatory information

Authorization status: Research-use only; no regulatory clearance stated.

Reproducibility

Baseline detection model per prior work; generators trained for 100 epochs with batch size 96 on NVIDIA A100 GPUs using PyTorch; bootstrap with 2000 resamples for AUC CIs; detailed training procedures provided in appendices.

Use

Intended: Detection and diagnosis

Out-of-scope: Other

Excluded: Diagnosis

User

Intended: Radiologist, Researcher

Out-of-scope: Layperson

Excluded: Patient