Development of a quantitative multiparametric ultrasound and deep learning classifier for the detection of prostate cancer.
[OBJECTIVES] Prostate cancer (PCa) diagnosis is increasingly guided by imaging, with ultrasound (US) emerging as a cost-effective and widely accessible modality.
- 95% CI 0.85-0.89
APA
Delberghe F, Li X, et al. (2026). Development of a quantitative multiparametric ultrasound and deep learning classifier for the detection of prostate cancer.. European radiology. https://doi.org/10.1007/s00330-026-12323-y
MLA
Delberghe F, et al.. "Development of a quantitative multiparametric ultrasound and deep learning classifier for the detection of prostate cancer.." European radiology, 2026.
PMID
41612079
Abstract
[OBJECTIVES] Prostate cancer (PCa) diagnosis is increasingly guided by imaging, with ultrasound (US) emerging as a cost-effective and widely accessible modality. This study develops a deep learning-based classifier predicting the presence of clinically significant (cs)PCa using quantitative features extracted from 3D multiparametric (mp)US.
[MATERIALS AND METHODS] A multicenter prospective cohort of 327 patients with suspicion of PCa underwent transrectal 3D mpUS scanning, including dynamic contrast-enhanced US and shear-wave elastography. Acquisitions were registered to 3D histology from radical prostatectomy, which served as the reference standard for the presence of csPCa. Voxels within lesions with International Society of Urological Pathology (ISUP) Grade Group ≥ 2 were considered malignant, and the rest were benign. A 3D deep learning classifier was trained on quantitative mpUS features to detect csPCa. The classifier was trained and internally evaluated on 250 patients and externally evaluated on 77 patients acquired later. Classifier performance was evaluated per voxel using the area under the receiver operating characteristic curve (ROC AUC).
[RESULTS] Using quantitative mpUS features from 327 patients, the classifier achieved a ROC AUC of 0.87 (95% CI: 0.85-0.89) on the internal evaluation set, using 7-fold cross-validation. On the external evaluation cohort, the classifier achieved a ROC AUC of 0.88 (95% CI: 0.87-0.89).
[CONCLUSION] The proposed classifier accurately detects csPCa using quantitative features from 3D mpUS and generalizes well to the external dataset. These results support mpUS as a promising, cost-effective tool for csPCa diagnosis.
[KEY POINTS] Question: Can quantitative features extracted from 3D multiparametric ultrasound (mpUS) reliably detect clinically significant prostate cancer (csPCa), enabling more accessible and affordable diagnosis?
[FINDINGS] Predicting csPCa using quantitative multiparametric ultrasound features achieved an area under the receiver operating characteristic curve of 0.87, increasing to 0.88 when externally evaluated.
[CLINICAL RELEVANCE] Our proposed deep learning-based classifier using quantitative 3D mpUS features accurately detects csPCa, as validated on the largest mpUS prostate dataset to date. This opens the door to ultrasound as an accurate, cost-effective method for csPCa detection.
[MATERIALS AND METHODS] A multicenter prospective cohort of 327 patients with suspicion of PCa underwent transrectal 3D mpUS scanning, including dynamic contrast-enhanced US and shear-wave elastography. Acquisitions were registered to 3D histology from radical prostatectomy, which served as the reference standard for the presence of csPCa. Voxels within lesions with International Society of Urological Pathology (ISUP) Grade Group ≥ 2 were considered malignant, and the rest were benign. A 3D deep learning classifier was trained on quantitative mpUS features to detect csPCa. The classifier was trained and internally evaluated on 250 patients and externally evaluated on 77 patients acquired later. Classifier performance was evaluated per voxel using the area under the receiver operating characteristic curve (ROC AUC).
[RESULTS] Using quantitative mpUS features from 327 patients, the classifier achieved a ROC AUC of 0.87 (95% CI: 0.85-0.89) on the internal evaluation set, using 7-fold cross-validation. On the external evaluation cohort, the classifier achieved a ROC AUC of 0.88 (95% CI: 0.87-0.89).
[CONCLUSION] The proposed classifier accurately detects csPCa using quantitative features from 3D mpUS and generalizes well to the external dataset. These results support mpUS as a promising, cost-effective tool for csPCa diagnosis.
[KEY POINTS] Question: Can quantitative features extracted from 3D multiparametric ultrasound (mpUS) reliably detect clinically significant prostate cancer (csPCa), enabling more accessible and affordable diagnosis?
[FINDINGS] Predicting csPCa using quantitative multiparametric ultrasound features achieved an area under the receiver operating characteristic curve of 0.87, increasing to 0.88 when externally evaluated.
[CLINICAL RELEVANCE] Our proposed deep learning-based classifier using quantitative 3D mpUS features accurately detects csPCa, as validated on the largest mpUS prostate dataset to date. This opens the door to ultrasound as an accurate, cost-effective method for csPCa detection.