MUSegNet: Automated detection of clinically significant prostate cancer on micro-ultrasound.
Authors
Affiliations (12)
Affiliations (12)
- Department of Radiology, Stanford University, Stanford, CA, USA. [email protected].
- Department of Urology, Stanford University, Stanford, CA, USA. [email protected].
- Department of Biomedical Data Science, Stanford University, Stanford, CA, USA. [email protected].
- Department of Urology, Stanford University, Stanford, CA, USA.
- Department of Radiology, Stanford University, Stanford, CA, USA.
- Department of Radiology, College of Medicine, The Catholic University of Korea, Seoul, Republic of Korea.
- Department of Urology, University of Alberta, Edmonton, AL, Canada.
- Department of Urology, University of California Los Angeles, Los Angeles, CA, USA.
- Department of Biomedical Sciences, Humanitas University, Pieve Emanuele, Milan, Italy.
- Department of Urology, IRCCS Humanitas Research Hospital, Milan, Italy.
- Department of Radiology, University of California San Diego, La Jolla, CA, USA.
- Department of Medicine, University of Florida, Gainesville, FL, USA.
Abstract
Micro-ultrasound (microUS) is a non-inferior alternative to multi-parametric MRI for clinically significant prostate cancer (csPCa) diagnosis, yet cancer's subtle features result in missed cancers and large inter-reader variability. We seek to automate the detection of csPCa on microUS by developing the microUS Prostate Cancer Detection Network (MUSegNet). Implemented using the well-established segmentation framework, nnUnet, MUSegNet automatically outlines the prostate, transitional zone, and csPCa. Our study explored strategies to train such segmentation models given reduced data, which is common with new imaging modalities such as microUS and/or when labeling data is tedious. MUSegNet was trained using three datasets: (1) 53 microUS exams manually labeled using MRI and pathology references (53 patients), (2) 565 unlabeled microUS exams (106 patients), and (3) 2749 conventional b-mode ultrasound exams (2281 patients). We evaluated MUSegNet in 73 independent patients from two institutions, against manual labels and in direct comparison with six expert readers. MUSegNet was significantly better at detecting csPCa than the baseline models trained using manual labels (area under the receiver operating characteristic curve - AUROC; 0.85 vs. 0.73, Delong p-value < 0.01, sensitivity: 0.74 vs. 0.27, dice similarity coefficient: 0.29 vs. 0.09, Wilcoxon p-value < 0.001). MUSegNet was better at detecting posterior vs. anterior lesions (AUROC: 0.94 vs. 0.72, sensitivity: 0.86 vs. 0.49, Dice: 0.31 vs. 0.11) and outperformed the six human readers for all lesions (AUROC: 0.85 vs. 0.78, sensitivity: 0.74 vs. 0.63). While evaluation in a larger cohort is needed, MUSegNet found csPCA in 33% of cases with csPCa and MRI invisible cancers (1 in 3). MUSegNet reliably detected the extent of clinically significant prostate cancer on micro-ultrasound exams, paving the way for improving cancer detection for biopsy targeting and reducing inter-reader variability. MUSegNet was best at capturing posterior cancers, with some benefit for the challenging anterior cancers.