Overview

Schema Version

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

Name

Report-guided Semi-supervised Learning (RG-SSL) for clinically significant prostate cancer detection on biparametric MRI

Link

https://fastmri.eu/research/bosma22a

Indexing

Keywords: Semisupervised learning, Pseudo labels, Report-guided SSL, Prostate cancer, Biparametric MRI, Annotation efficiency, Computer-aided detection, nnU-Net

Content: GU, MR, OI

RadLex: RID348, RID50295, RID50316, RID50313, RID50294, RID50301

Author(s)

Joeran S. Bosma

Anindo Saha

Matin Hosseinzadeh

Ivan Slootweg

Maarten de Rooij

Henkjan Huisman

Organization(s)

Diagnostic Image Analysis Group, Department of Medical Imaging, Radboud University Medical Center

Ziekenhuisgroep Twente

Version

1.0

License

Text: CC BY 4.0

URL: https://creativecommons.org/licenses/by/4.0/

Funding

Supported in part by Health~Holland (LSHM20103), European Union H2020 ProCAncer-I project (952159), European Union H2020 PANCAIM project (101016851), and Siemens Healthineers (CID: C00225450).

Ethical review

Retrospective study; written informed consent was waived by the institutional review board.

Date

Updated: 2023-06-22

Published: 2023-07-26

Created: 2023-02-04

References

[1] Bosma JS, Saha A, Hosseinzadeh M, Slootweg I, de Rooij M, Huisman H. "Semisupervised Learning with Report-guided Pseudo Labels for Deep Learning–based Prostate Cancer Detection Using Biparametric MRI". Radiology: Artificial Intelligence. 2023 Sep;5(5):e230031.. 2023-07-26. doi:10.1148/ryai.230031. PMID: 37795142. PMCID: PMC10546362.

Model

Architecture

Voxel-level segmentation using nnU-Net (self-configuring 3D CNN) with ensemble teacher models to generate pseudo labels; student model trained on manual + pseudo labels.

Availability

Code publicly available: https://fastmri.eu/research/bosma22a

Clinical benefit

Assists detection and risk stratification of clinically significant prostate cancer (GGG ≥ 2) on bpMRI; achieves SL-comparable performance with substantially fewer manual annotations.

Clinical workflow phase

Clinical decision support systems; workflow optimization (reducing annotation burden in model development).

Degree of automation

Automates lesion candidate generation and segmentation during training via pseudo labels; outputs lesion detections/segmentations for reader support.

Indications for use

Development and evaluation focused on detection of clinically significant prostate cancer in patients undergoing prostate bpMRI in hospital radiology settings.

Input

Biparametric prostate MRI: axial T2-weighted images, calculated high-b-value DWI (≥1400 s/mm^2), and ADC maps.

Instructions

Train teacher on manually labeled subset; extract number of clinically significant lesions (PI-RADS ≥4) per exam from diagnostic reports; generate pseudo labels by keeping n_sig highest-confidence lesion candidates from teacher heatmaps; train student on combined manual+pseudo labels; optional second iteration using student as teacher.

Limitations

Potential for reinforcing incorrect model predictions with pseudo labels; report parsing is rule-based and may require adaptation to different report structures/languages; training labels based on PI-RADS ≥4 exclude PI-RADS 3, which may bias models; pseudo labels exclude exams with fewer detected candidates than reported lesions.

Output

CDEs: RDE1558, RDE1554, RDE1550

Description: Voxel-level segmentation masks of suspected clinically significant prostate cancer lesions, lesion candidates for detection, and examination-level diagnostic likelihood/performance.

Reproducibility

Models trained with fivefold cross-validation and multiple restarts (RG-SSL: 3 restarts; baseline SSL: 1–2 restarts); performance reported across independent runs with permutation testing and bootstrapping for CIs.

Use

Intended: Detection and diagnosis

Out-of-scope: Report translation, Detection

Excluded: Diagnosis

User

Intended: Radiologist, Researcher

Out-of-scope: Layperson