Related Experiment Video
Updated: Sep 4, 2025
A Cognitive Fusion-guided Prostate Biopsy Using Multiparametric Magnetic Resonance Imaging and Transrectal Ultrasound
Published on: March 21, 2025
Prostate158 - An expert-annotated 3T MRI dataset and algorithm for prostate cancer detection
Lisa C Adams1, Marcus R Makowski2, Günther Engel3
1Charité - Universitätsmedizin Berlin, Corporate Member of Freie Universität Berlin and Humboldt Universität zu Berlin, Institute for Radiology, Luisenstraße 7, 10117, Hindenburgdamm 30, 12203, Berlin, Germany; Berlin Institute of Health at Charité - Universitätsmedizin Berlin, Charitéplatz 1, 10117, Berlin, Germany.
This study introduces Prostate158, an expert-annotated prostate MRI dataset and benchmark, to advance deep learning for prostate cancer segmentation. The developed U-ResNet models provide reliable baseline performance for reproducible research.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Deep learning (DL) for prostate segmentation on MRI requires expert-annotated data and baselines, which are often unavailable.
- Lack of public data and benchmarks hinders reproducibility and comparability in prostate cancer research.
Purpose of the Study:
- To introduce Prostate158, a novel, expert-annotated dataset of 158 biparametric 3T prostate MRIs.
- To establish reliable baseline algorithms (U-ResNets) for prostate anatomy and cancer lesion segmentation.
- To foster reproducibility and comparability in developing DL models for prostate MRI segmentation.
Main Methods:
- Dataset creation: 158 expert-annotated biparametric 3T prostate MRIs (T2w, ADC maps).
- Algorithm development: Two U-ResNets trained for segmenting central gland, peripheral zone, and PI-RADS ≥4 lesions.
- Performance evaluation: Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), Average Surface Distance (ASD); Wilcoxon test; assessment on external datasets (Medical Segmentation Decathlon, PROSTATEx).
Main Results:
- Baseline U-ResNets achieved high segmentation performance, comparable to interrater agreement for central gland and peripheral zone.
- Segmentation performance for PCa lesions (DSC) was 0.45 (Reader 1) and 0.40 (Reader 2), with interrater agreement at 0.6.
- The model demonstrated generalizability on external datasets, with DSCs of 0.82 and 0.86 for the central gland, and 0.64 and 0.71 for the peripheral zone.
Conclusions:
- Prostate158 provides an openly accessible, expert-annotated dataset for prostate MRI.
- The established U-ResNet benchmarks facilitate reproducible research and development of prostate segmentation algorithms.
- This resource aims to accelerate advancements in AI-driven prostate cancer diagnosis and management.
