Related Experiment Video
Updated: Jun 16, 2025
A Cognitive Fusion-guided Prostate Biopsy Using Multiparametric Magnetic Resonance Imaging and Transrectal Ultrasound
Published on: March 21, 2025
Deep Learning-based Unsupervised Domain Adaptation via a Unified Model for Prostate Lesion Detection Using Multisite
Hao Li1, Han Liu1, Heinrich von Busch1
1From Digital Technology and Innovation, Siemens Healthineers, 755 College Rd E, Princeton, NJ 08540 (H. Li, H. Liu, D.C., A.K., B.L.); Diagnostic Imaging, Siemens Healthineers, Erlangen, Bavaria, Germany (H.v.B., R.G.); Vanderbilt University, Nashville, Tenn (H. Li, H. Liu, I.O.); Radboud University Medical Center, Nijmegen, the Netherlands (H.H.); New York University, New York, NY (A.T.); Universitätsspital Basel, Basel, Switzerland (D.W.); Charité, Universitätsmedizin Berlin, Berlin, Germany (T.P.); Patero Clinic, Moscow, Russia (I.S.); Eunpyeong St. Mary's Hospital, Catholic University of Korea, Seoul, Republic of Korea (M.H.C.); Department of Radiology, Changhai Hospital of Shanghai, Shanghai, China (Q.Y.); Diagnostikum Graz Süd-West, Graz, Austria (D.S.); Department of Radiology, Loyola University Medical Center, Maywood, Ill (S.S.); Department of Diagnostic Radiology, Oregon Health and Science University School of Medicine, Portland, Ore (F.C.); and Massachusetts General Hospital, Boston, Mass (M.H.).
Unsupervised domain adaptation using generated images significantly improves prostate cancer detection accuracy in multisite MRI scans. This method enhances supervised learning models, especially with non-standard imaging protocols.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Medicine
- Oncology
Background:
- Prostate cancer (PCa) detection using biparametric MRI (bpMRI) faces challenges due to variations in imaging protocols across different centers.
- Supervised learning (SL) models for PCa detection can be limited by domain shift in multisite datasets.
- Standardizing image acquisition parameters, particularly b values in diffusion-weighted imaging (DWI), is crucial for reliable PCa detection.
Purpose of the Study:
- To evaluate the efficacy of an unsupervised domain adaptation (UDA) method employing generated images to enhance SL model performance for PCa detection.
- To assess the impact of UDA on multisite bpMRI datasets with varying DWI acquisition parameters.
- To determine if UDA can mitigate performance degradation caused by deviations from PI-RADS-recommended imaging protocols.
Main Methods:
- A novel UDA method was developed using a unified generative model to translate DWI (ADC and DW images) across different b values.
- Generated images were used to replace original images, aligning them with PI-RADS-recommended b values for PCa detection.
- The study utilized a large retrospective dataset from 5150 patients across nine imaging centers, with an independent test set of 1692 cases for evaluation.
Main Results:
- The UDA method significantly improved the area under the receiver operating characteristic curve (AUC) for PCa lesion detection compared to baseline SL.
- Overall AUC increased from 0.73 to 0.79 for PI-RADS 3+ lesions and 0.77 to 0.80 for PI-RADS 4+ lesions.
- In challenging cases with significant protocol deviations, UDA boosted AUC from 0.49 to 0.76 (PI-RADS 3+) and 0.50 to 0.77 (PI-RADS 4+).
Conclusions:
- Unsupervised domain adaptation with generated images effectively enhances SL model performance for PCa detection in multisite bpMRI.
- The UDA approach demonstrates particular benefit in improving detection accuracy for datasets acquired with non-standard DWI protocols.
- This method offers a promising strategy to improve the generalizability and robustness of AI-based PCa detection systems across diverse clinical settings.
