Artificial intelligence for prostate cancer detection and risk stratification using transrectal ultrasound: a narrative review

Introduction

Prostate cancer (PCa) remains a leading cause of cancer morbidity and mortality globally (1,2). Unlike many other malignant tumors that progress rapidly, early-stage clinically insignificant PCa (CISPCa) is characterized by low risk and slow progression, and is managed primarily with active surveillance rather than immediate treatment. In contrast, intermediate-to-advanced clinically significant PCa (CSPCa) carries a high risk of progression and a poor prognosis, requiring aggressive intervention (3-5). Therefore, clinical practice prioritizes identifying CSPCa and accurately stratifying PCa risk. Pathological assessment based on the Gleason score (GS) remains the gold standard for diagnosing PCa and guiding risk stratification (6). Clinically, a GS of ≥3+3 confirms PCa (6), with subsequent stratification into CSPCa (GS >3+3) and CISPCa (GS ≤3+3) based on pathological features.

PCa risk stratification relies on two primary frameworks:

• The D’Amico risk stratification, adopted by the European Association of Urology (EAU), stratifies PCa into three categories: low-risk (GS ≤6), intermediate-risk (GS =7), and high-risk (GS ≥8) (7,8) (hereafter referred to as the “three-tier risk stratification”).
• The Grade Group (GG) system proposed by the International Society of Urological Pathology (ISUP) (6), which refines risk stratification into five tiers: GG1 (GS 3+3), GG2 (GS 3+4), GG3 (GS 4+3), GG4 (GS 8), and GG5 (GS 9 or 10) (hereafter referred to as the “five-tier risk stratification”).

For this review, PCa was further subdivided into subgroups based on these two systems:

• Based on the three-tier risk stratification, five subgroups were defined: low-risk (GS ≤6), intermediate-risk (GS =7), high-risk (GS ≥8), low-to-intermediate risk (GS ≤7), and intermediate-to-high risk (GS ≥7).
• Based on the five-tier risk stratification, 11 subgroups were defined: low-risk (GG1), low-to-intermediate risk (GG2), intermediate-risk (GG3), intermediate-to-high risk (GG4), high-risk (GG5), below-intermediate risk (GG1–2), intermediate-risk and below (GG1–3), below-high risk (GG1–4), above-low risk (GG2–5), intermediate-risk and above (GG3–5), and above-intermediate risk (GG4–5).

Currently, multiparametric magnetic resonance imaging (mpMRI) serves as the reference standard for PCa detection and local staging (9). However, its widespread adoption is hindered by high costs and limited accessibility. In contrast, transrectal ultrasound (TRUS) remains the most prevalent imaging modality due to its cost-effectiveness, real-time capabilities, and integral role in biopsy guidance (3,10,11). Notwithstanding these advantages, its limited resolution and high interobserver variability impede accurate risk assessment. Integrating artificial intelligence (AI) into TRUS workflows, therefore, presents a critical opportunity to enhance diagnostic precision, offering a viable alternative to resource-intensive mpMRI and addressing an unmet need in urological practice.

To our knowledge, existing research on TRUS-based AI has primarily focused on PCa detection and diagnosis (12-15). While several studies have explored the utility of AI for CSPCa identification via TRUS, reporting area under the curve (AUC) values of 0.85–0.89 (16), its application in risk stratification remains inadequately explored. Notably, only one study has investigated AI-based prediction of three-tier risk stratification, achieving a diagnostic accuracy of 0.830 (7), and comprehensive research on AI-driven differentiation across all five ISUP GG tiers is currently lacking. This review summarizes the current status and recent advances in TRUS-based AI models for PCa diagnosis, CSPCa identification, and risk stratification. We present this article in accordance with the Narrative Review reporting checklist (available at https://gs.amegroups.com/article/view/10.21037/gs-2025-aw-504/rc).

Methods

We systematically searched the PubMed and Web of Science databases for original research articles published from 1999 to 2026, focusing on AI applications in ultrasound for PCa detection. Only peer-reviewed, English-language studies were included. The detailed search strategy, including keywords and MeSH terms, is outlined in Table 1.

Discussion

Table 2 summarizes a subgroup analysis of AI-based ultrasound modalities for PCa, comparing the key technical features of distinct TRUS-based models. TRUS, centered on two/three-dimensional (2D/3D) deep learning, offers accessible data (AUC: 0.78–0.96) (16,21,22,27), broad compatibility with standard equipment, and high clinical utility; however, 2D models are vulnerable to image artifacts, whereas 3D models demand substantial computing power, necessitating lightweight optimization and small-sample expansion strategies. Temporal enhanced ultrasound (TeUS) captures dynamic features via recurrent neural networks (RNNs) or unsupervised deep neural networks, achieving an AUC >0.8 for high-grade PCa (14); yet, its reliance on specialized equipment, time-consuming sequence processing, and limited sample sizes (<160 cases) (12,14,29) hinders scalability, posing challenges for multi-center validation and real-time implementation. Contrast-enhanced ultrasound (CEUS) utilizes 3D convolutional neural networks (CNNs) or multi-modal fusion to localize lesions (specificity: 82–91.45%), which is particularly applicable to MRI-negative patients (30), but is constrained by the high cost of targeted contrast agents and operational complexity. Multiparametric ultrasound (mpUS) integrates B-mode, shear wave elastography (SWE), and CEUS, demonstrating an AUC of 0.90 for PCa with GS >3+4 (31), though it suffers from cumbersome data acquisition and significant multi-modal noise. Finally, micro-ultrasound identifies small lesions via high resolution and self-supervised learning (AUC, 91%) (36), but its utility is limited by high equipment costs, restricted availability, and insufficient clinical validation data.

Notably, few systematic reviews synthesize TRUS-based AI models for PCa detection, CSPCa identification, and PCa risk stratification. The current research status in this field is summarized below:

Current status of AI-based PCa diagnosis using TRUS images

In 2019, Feng et al. conducted an experiment using 20 mice implanted with human PCa xenografts. They acquired CEUS videos of the tumor lesions and constructed a 3D CNN model for PCa detection, leveraging sequential CEUS frames (30). Specifically, the model performed 3D convolutional operations to uniformly extract spatial and temporal features from sequential CEUS frames, enabling automated PCa detection in ultrasound image sequences. Notably, this study attempted to integrate CEUS with a prostate-specific membrane antigen (PSMA)-targeted contrast agent for automated PCa detection and compared this approach with non-targeted blank agents. Results showed that the targeted deep learning model improved PCa detection sensitivity by >10% compared with non-targeted agents. For targeted CEUS images, the model achieved a specificity of >91% and accuracy of 90%.

In 2020, Huang et al. used biopsy results as the reference standard to collect longitudinal prostate images acquired before and during biopsy (48 patients, 342 pairs of images) (20). Based on the positional correspondence of the biopsy needle between pre- and intra-biopsy images, researchers manually annotated needle positions from intra-biopsy images onto pre-biopsy images. Rectangular image strips in pre-biopsy images—corresponding to biopsy needle positions—were designated as regions of interest (ROIs). Texture features were extracted from preprocessed ROIs, linearly combined, and input into a support vector machine (SVM) classifier for PCa classification. This approach achieved a diagnostic accuracy of 70.93%, sensitivity of 70.00%, and specificity of 71.74% (20). Due to the small sample size, the authors did not explore the model’s ability to distinguish between three-tier or five-tier PCa risk categories.

In 2021, Liu et al. similarly used biopsy results as the reference standard to collect 702 cross-sectional TRUS prostate images (21). First, three specialists with >5 years of clinical experience delineated prostate tissue boundaries to establish image labels. A deep learning model combining S-Mask region-based convolutional neural network (R-CNN) and Inception-v3 was developed for auxiliary PCa diagnosis. The modified S-Mask R-CNN enabled accurate segmentation of prostate ultrasound images and generation of candidate regions, with an ROI alignment algorithm for pixel-level feature localization. Convolutional networks generated binary masks to segment prostate tissue from the background, and the Inception-v3 network was used for lesion detection. Experimental results showed this method achieved an 80% accuracy for PCa detection.

In 2022, Hassan et al. proposed a novel automated classification algorithm for PCa detection from TRUS and MRI images by fusing multiple deep learning approaches (15). Custom layers were added to pre-trained models (MobileNetV2, InceptionResNetV2, ResNet50V2, ResNet101V2, ResNet152V2, Xception, VGG16, VGG19, and InceptionV3), which were applied to MRI and TRUS datasets from The Cancer Imaging Archive (611,119 images). The best-performing model achieved 97% accuracy on TRUS test data and 80% on MRI test data. To improve MRI performance, the top model was used as a feature extractor and combined with shallow machine learning algorithms (SVM, gradient boosting, and random forest) to develop a fused model. This fusion strategy significantly enhanced PCa detection, increasing MRI accuracy from 80% to 88%.

In 2024, Li et al. collected 1,696 pairs of pre- and intra-biopsy oblique-section TRUS grayscale images from 142 patients. Using biopsy results as the reference standard, they trained a ResNet50 model with three input modalities: original images (Whole model), biopsy needle tract images (Needle model), and combined images [feature pyramid networks (FPN)] (25). For the Needle model, all images underwent bilinear interpolation followed by symmetric normalization (a widely used medical image registration algorithm). The resulting deformation field was interpolated onto intra-biopsy images to generate registered images. In the test set, the FPN achieved an AUC of 0.934, specificity of 0.966, and sensitivity of 0.829—significantly outperforming the Whole model (AUC =0.908, specificity =0.938, sensitivity =0.754) and Needle model (AUC =0.905, specificity =0.918, sensitivity =0.764) (P<0.05) (25). However, the small sample size (1,696 2D TRUS images from 142 patients) precluded exploration of the model’s ability to stratify PCa into three-tier or five-tier risk groups.

Current status of AI-based CSPCa diagnosis using TRUS images

In 2023, Sun et al. prospectively evaluated pre-biopsy cross-sectional TRUS videos from 832 patients across four centers to develop an AI model for CSPCa identification (16). Using prostate biopsy or radical prostatectomy pathology as the reference standard, they constructed 2D and 3D P-Net CNN models using a training dataset (559 patients, with only intra-prostate contour images included). Models were tested on internal (140 patients) and external (133 patients) validation datasets. For CSPCa prediction, the 3D P-Net achieved an AUC of 0.89 [95% confidence interval (CI): 0.83–0.95] in internal testing and 0.85 (95% CI: 0.78–0.93) in external validation, whereas the 2D P-Net yielded AUCs of 0.86 (95% CI: 0.80–0.93) and 0.79 (95% CI: 0.71–0.87), respectively (16). Similar to prior studies, the authors did not investigate the models’ performance in three-tier or five-tier PCa risk stratification.

In 2023, Zhang et al. explored the value of a machine learning model integrating multimodal TRUS (grayscale, real-time elastography, CEUS) and prostate-specific antigen (PSA)-related parameters (total serum PSA, free PSA, free-to-total PSA ratio) for CSPCa diagnosis (32). Pathological results from 301 biopsy patients (218 CSPCa, 83 CISPCa) served as the gold standard. Six machine learning models were compared: artificial neural network (ANN), logistic regression, SVM, decision tree, random forest, and k-nearest neighbor. The ANN model performed best, achieving a sensitivity of 80%, specificity of 88.6%, and AUC of 0.855 (32).

In 2024, Wu et al. collected cross-sectional TRUS videos from 512 biopsy patients (346 CSPCa, 166 CISPCa), all of whom underwent SWE (34). A CSPCa identification model was developed using these multimodal data: two 3D ResNet-50 models extracted features from TRUS grayscale hypoechoic regions and SWE abnormally stiff regions, respectively. An adaptive spatial fusion module aggregated features from the two modalities, with orthogonal regularization to reduce feature redundancy and enhance weight orthogonality. The model achieved a favorable AUC of 0.84 for CSPCa identification.

In 2024, Takeda et al. collected multimodal data from 151 ultrasound-guided biopsy patients, including PSA results, TRUS images, and MRI sequences [T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and apparent diffusion coefficient (ADC) maps] (38). A pre-trained Xception CNN (39) was evaluated for CSPCa prediction in the full cohort (n=151) and a subset with PSA ≤20 ng/mL (n=122). AUC values for CSPCa prediction in the full cohort were as follows: (I) PSA testing: 0.649 (95% CI: 0.467–0.832); (II) TRUS images: 0.715 (95% CI: 0.551–0.878); (III) T2WI: 0.738 (95% CI: 0.581–0.895); (IV) DWI: 0.582 (95% CI: 0.396–0.767); (V) ADC maps: 0.690 (95% CI: 0.519–0.861); (VI) multimodal data: 0.878 (95% CI: 0.772–0.984). For the PSA ≤20 ng/mL subset, AUC values were: (I) PSA testing: 0.574 (95% CI: 0.330–0.819); (II) TRUS images: 0.708 (95% CI: 0.508–0.908); (III) T2WI: 0.803 (95% CI: 0.629–0.976); (IV) DWI: 0.564 (95% CI: 0.341–0.787); (V) ADC maps: 0.662 (95% CI: 0.449–0.874); (VI) multimodal data: 0.862 (95% CI: 0.723–1.000).

Current status of AI-based three-tier PCa risk stratification using TRUS images

The EAU adopts the D’Amico risk stratification, stratifying PCa into low-risk (GS ≤6), intermediate-risk (GS =7), and high-risk (GS ≥8) (7,8). Based on this system, PCa can be further subdivided into five subgroups: low-risk (GS ≤6), intermediate-risk (GS =7), high-risk (GS ≥8), low-to-intermediate risk (GS ≤7), and intermediate-to-high risk (GS ≥7).

AI-based prediction of low-to-intermediate-risk (GS ≤7) vs. high-risk (GS ≥8) PCa

In 2022, Akatsuka et al. tested three pre-trained deep CNN models (Xception, InceptionV3, VGG16) and selected Xception for its superior performance in ultrasound image classification (22). A total of 2,676 cross-sectional TRUS images from 691 patients were divided into training and test sets based on ultrasound evaluation date. Two deep learning analyses were performed with different labeling strategies: first, distinguishing PCa (GS ≥6) from non-PCa (AUC =0.693); second, distinguishing high-risk (GS ≥8) from low-to-intermediate-risk (GS ≤7) PCa (AUC =0.723) (22). Integrating TRUS images with clinical data (age, PSA levels) into the Xception model substantially improved high-risk PCa prediction (AUC =0.835) (22). Research in this domain remains limited, and these findings require validation across different ultrasound devices and observer cohorts.

AI-based prediction of low-risk (GS ≤6), intermediate-risk (GS =7), and high-risk (GS ≥8) PCa

In 2022, Lu et al. collected 525 cross-sectional TRUS images of biopsy-confirmed PCa, integrating biopsy needle tract information, biopsy results, and MRI findings (7). PCa lesions were manually delineated to label ROIs, and a deep learning-based Automatic Region-based Gleason Grading network was proposed to differentiate low-, intermediate-, and high-risk PCa. The network first identified PCa lesions via a region-marked object detection network, followed by Gleason grading of marked regions. The model achieved an accuracy of 0.830 for three-tier risk stratification. However, the authors did not investigate the model’s utility for five-tier GG stratification or its performance using TRUS alone (without integrating MRI or biopsy data).

Current status of AI-based five-tier PCa risk stratification using TRUS images

With the growing demand for refined risk stratification, the ISUP classified PCa into five GGs (five-tier risk stratification) (6): GG1 (GS 3+3), GG2 (GS 3+4), GG3 (GS 4+3), GG4 (GS 8), and GG5 (GS 9 or 10). Based on this system, PCa can be subdivided into 11 subgroups as outlined in the Introduction section.

AI-based prediction of below-intermediate-risk (GG1–2) vs. intermediate-risk and above (GG3–5) PCa

In 2023, Lorusso et al. retrospectively analyzed cross-sectional TRUS images from 64 patients, using post-radical prostatectomy whole-mount pathology as the reference standard (24). A computer-aided ANN analysis model was proposed, and suspicious lesions identified by the model were compared with pathological cancer foci. For detecting intermediate-risk and above PCa (GG3–5), the model achieved a sensitivity of 69%, specificity of 77%, negative predictive value of 88%, positive predictive value of 50%, and accuracy of 75%. Tumor volume did not affect the model’s diagnostic performance (24).

In 2020, Wildeboer et al. collected TRUS grayscale images, SWE data, and CEUS data from 50 biopsy-confirmed PCa patients, using radical prostatectomy pathology as the reference standard (31). A random forest classifier was developed based on multimodal TRUS, involving four steps: prostate segmentation [via a deep learning-based algorithm (40)], data registration, feature extraction, and multiparametric classification. Hemodynamic features from CEUS videos were extracted and converted to 2D to match SWE and grayscale image dimensions. Radiomic features (pixel values, lesion capsule spiculation, perfusion/elasticity asymmetry) were extracted from hemodynamic maps, SWE, and grayscale images, then combined via random forest. The classifier achieved an AUC of 0.75 for PCa diagnosis and 0.90 for predicting intermediate-risk and above PCa (GG3–5) (31).

AI-based prediction of below-high-risk (GG1–4) vs. high-risk (GG5) PCa

To date, no literature has reported AI-based TRUS models for binary classification of below-high-risk (GG1–4) vs. high-risk (GG5) PCa.

AI-based prediction of all five GG tiers (GG1–GG5) PCa

To date, no studies have reported AI-based TRUS models for five-class PCa classification into GG1 (low-risk), GG2 (low-to-intermediate-risk), GG3 (intermediate-risk), GG4 (intermediate-to-high-risk), and GG5 (high-risk) groups.

Conclusions

Overall, TRUS-based AI for PCa risk stratification (three-tier and five-tier) remains underdeveloped, with a paucity of relevant studies. Key challenges are summarized as follows: (I) Unclear PCa lesion localization in some TRUS images hinders accurate annotation for model training. (II) Assigning risk stratification labels to TRUS lesions requires specialized clinical and pathological expertise, increasing dataset annotation difficulty. (III) Insufficient data for specific risk groups and imbalanced group distribution are common, leading to model overfitting, poor generalization, and limited utility for refined stratification (e.g., five-tier). (IV) Most studies rely on single-center retrospective data, and their generalizability requires validation in multi-center prospective cohorts.

Future advancements in TRUS equipment, novel technologies, high-quality dataset accumulation, AI algorithm optimization, multimodal data integration, and multi-institutional collaboration are expected to enable TRUS-based AI to accurately predict and diagnose PCa, CSPCa, and both three-tier (EAU) and five-tier (ISUP) risk stratification in real time. Such progress will provide reliable evidence for treatment decision-making, improving the efficiency of PCa diagnosis and management.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the Narrative Review reporting checklist. Available at https://gs.amegroups.com/article/view/10.21037/gs-2025-aw-504/rc

Peer Review File: Available at https://gs.amegroups.com/article/view/10.21037/gs-2025-aw-504/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://gs.amegroups.com/article/view/10.21037/gs-2025-aw-504/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Rawla P. Epidemiology of Prostate Cancer. World J Oncol 2019;10:63-89. [Crossref] [PubMed]
2. Culp MB, Soerjomataram I, Efstathiou JA, et al. Recent Global Patterns in Prostate Cancer Incidence and Mortality Rates. Eur Urol 2020;77:38-52. [Crossref] [PubMed]
3. Mottet N, Bellmunt J, Bolla M, et al. EAU-ESTRO-SIOG Guidelines on Prostate Cancer. Part 1: Screening, Diagnosis, and Local Treatment with Curative Intent. Eur Urol 2017;71:618-29. [Crossref] [PubMed]
4. van der Leest M, Cornel E, Israël B, et al. Head-to-head Comparison of Transrectal Ultrasound-guided Prostate Biopsy Versus Multiparametric Prostate Resonance Imaging with Subsequent Magnetic Resonance-guided Biopsy in Biopsy-naïve Men with Elevated Prostate-specific Antigen: A Large Prospective Multicenter Clinical Study. Eur Urol 2019;75:570-8. [Crossref] [PubMed]
5. Briganti A, Fossati N, Catto JWF, et al. Active Surveillance for Low-risk Prostate Cancer: The European Association of Urology Position in 2018. Eur Urol 2018;74:357-68. [Crossref] [PubMed]
6. Epstein JI, Egevad L, Amin MB, et al. The 2014 International Society of Urological Pathology (ISUP) Consensus Conference on Gleason Grading of Prostatic Carcinoma: Definition of Grading Patterns and Proposal for a New Grading System. Am J Surg Pathol 2016;40:244-52. [Crossref] [PubMed]
7. Lu X, Zhang S, Liu Z, et al. Ultrasonographic pathological grading of prostate cancer using automatic region-based Gleason grading network. Comput Med Imaging Graph 2022;102:102125. [Crossref] [PubMed]
8. D'Amico AV, Whittington R, Malkowicz SB, et al. Biochemical outcome after radical prostatectomy, external beam radiation therapy, or interstitial radiation therapy for clinically localized prostate cancer. JAMA 1998;280:969-74. [Crossref] [PubMed]
9. Cornford P, van den Bergh RCN, Briers E, et al. EAU-EANM-ESTRO-ESUR-ISUP-SIOG Guidelines on Prostate Cancer-2024 Update. Part I: Screening, Diagnosis, and Local Treatment with Curative Intent. Eur Urol 2024;86:148-63. [Crossref] [PubMed]
10. Wu P, Liu Y, Li Y, et al. Robust Prostate Segmentation Using Intrinsic Properties of TRUS Images. IEEE Trans Med Imaging 2015;34:1321-35. [Crossref] [PubMed]
11. Heidenreich A, Abrahamsson PA, Artibani W, et al. Early detection of prostate cancer: European Association of Urology recommendation. Eur Urol 2013;64:347-54. [Crossref] [PubMed]
12. Azizi S, Bayat S, Yan P, et al. Deep Recurrent Neural Networks for Prostate Cancer Detection: Analysis of Temporal Enhanced Ultrasound. IEEE Trans Med Imaging 2018;37:2695-703. [Crossref] [PubMed]
13. Azizi S, Van Woudenberg N, Sojoudi S, et al. Toward a real-time system for temporal enhanced ultrasound-guided prostate biopsy. Int J Comput Assist Radiol Surg 2018;13:1201-9. [Crossref] [PubMed]
14. Sedghi A, Pesteie M, Javadi G, et al. Deep neural maps for unsupervised visualization of high-grade cancer in prostate biopsies. Int J Comput Assist Radiol Surg 2019;14:1009-16. [Crossref] [PubMed]
15. Hassan MR, Islam MF, Uddin MZ, et al. Prostate cancer classification from ultrasound and MRI images using deep learning based Explainable Artificial Intelligence. Future Generat Comput Syst 2022;127:462-72.
16. Sun YK, Zhou BY, Miao Y, et al. Three-dimensional convolutional neural network model to identify clinically significant prostate cancer in transrectal ultrasound videos: a prospective, multi-institutional, diagnostic study. EClinicalMedicine 2023;60:102027. [Crossref] [PubMed]
17. Ronco AL, Fernández R. Improving ultrasonographic diagnosis of prostate cancer with neural networks. Ultrasound Med Biol 1999;25:729-33. [Crossref] [PubMed]
18. Han SM, Lee HJ, Choi JY. Computer-aided prostate cancer detection using texture features and clinical features in ultrasound image. J Digit Imaging 2008;21:S121-33. [Crossref] [PubMed]
19. Glotsos D, Kalatzis I, Theocharakis P, et al. A multi-classifier system for the characterization of normal, infectious, and cancerous prostate tissues employing transrectal ultrasound images. Comput Methods Programs Biomed 2010;97:53-61. [Crossref] [PubMed]
20. Huang X, Chen M, Liu P, et al. Texture Feature-Based Classification on Transrectal Ultrasound Image for Prostatic Cancer Detection. Comput Math Methods Med 2020;2020:7359375. [Crossref] [PubMed]
21. Liu Z, Yang C, Huang J, et al. Deep learning framework based on integration of S-Mask R-CNN and Inception-v3 for ultrasound image-aided diagnosis of prostate cancer. Future Gener Comput Syst 2021;114:358-67.
22. Akatsuka J, Numata Y, Morikawa H, et al. A data-driven ultrasound approach discriminates pathological high grade prostate cancer. Sci Rep 2022;12:860. [Crossref] [PubMed]
23. Lu X, Liu X, Xiao Z, et al. Self-supervised dual-head attentional bootstrap learning network for prostate cancer screening in transrectal ultrasound images. Comput Biol Med 2023;165:107337. [Crossref] [PubMed]
24. Lorusso V, Kabre B, Pignot G, et al. External validation of the computerized analysis of TRUS of the prostate with the ANNA/C-TRUS system: a potential role of artificial intelligence for improving prostate cancer detection. World J Urol 2023;41:619-25. [Crossref] [PubMed]
25. Li S, Ye X, Tian H, et al. An artificial intelligence model based on transrectal ultrasound images of biopsy needle tract tissues to differentiate prostate cancer. Postgrad Med J 2024;100:228-36. [Crossref] [PubMed]
26. Lu X, Guo Y, Zhang S, et al. Deep Augmented Metric Learning Network for Prostate Cancer Classification in Ultrasound Images. IEEE J Biomed Health Inform 2025;29:1849-60. [Crossref] [PubMed]
27. Lou W, Chen P, Wu C, et al. Prostate cancer classification using 3D deep learning and ultrasound video clips: a multicenter study. Front Oncol 2025;15:1582035. [Crossref] [PubMed]
28. Rusu M, Jahanandish H, Vesal S, et al. ProCUSNet: Prostate Cancer Detection on B-mode Transrectal Ultrasound Using Artificial Intelligence for Targeting During Prostate Biopsies. Eur Urol Oncol 2025;8:477-85. [Crossref] [PubMed]
29. Javadi G, Samadi S, Bayat S, et al. Training deep neural networks with noisy clinical labels: toward accurate detection of prostate cancer in US data. Int J Comput Assist Radiol Surg 2022;17:1697-705. [Crossref] [PubMed]
30. Feng Y, Yang F, Zhou X, et al. A Deep Learning Approach for Targeted Contrast-Enhanced Ultrasound Based Prostate Cancer Detection. IEEE/ACM Trans Comput Biol Bioinform 2019;16:1794-801. [Crossref] [PubMed]
31. Wildeboer RR, Mannaerts CK, van Sloun RJG, et al. Automated multiparametric localization of prostate cancer based on B-mode, shear-wave elastography, and contrast-enhanced ultrasound radiomics. Eur Radiol 2020;30:806-15. [Crossref] [PubMed]
32. Zhang M, Liu Y, Yao J, et al. Value of machine learning-based transrectal multimodal ultrasound combined with PSA-related indicators in the diagnosis of clinically significant prostate cancer. Front Endocrinol (Lausanne) 2023;14:1137322. [Crossref] [PubMed]
33. Jager A, Postema AW, Mischi M, et al. Clinical Trial Protocol: Developing an Image Classification Algorithm for Prostate Cancer Diagnosis on Three-dimensional Multiparametric Transrectal Ultrasound. Eur Urol Open Sci 2023;49:32-43. [Crossref] [PubMed]
34. Wu H, Fu J, Ye H, et al. Multi-Modality Transrectal Ultrasound Video Classification for Identification of Clinically Significant Prostate Cancer. 2024 IEEE International Symposium on Biomedical Imaging (ISBI). IEEE; 2024:1-5.
35. Wu H, Fu J, Ye H, et al. Towards Multi-modality Fusion and Prototype-Based Feature Refinement for Clinically Significant Prostate Cancer Classification in Transrectal Ultrasound. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland; 2024:724-33.
36. Wilson PFR, Gilany M, Jamzad A, et al. Self-Supervised Learning With Limited Labeled Data for Prostate Cancer Detection in High-Frequency Ultrasound. IEEE Trans Ultrason Ferroelectr Freq Control 2023;70:1073-83. [Crossref] [PubMed]
37. Harmanani M, Wilson PFR, To MNN, et al. TRUSWorthy: toward clinically applicable deep learning for confident detection of prostate cancer in micro-ultrasound. Int J Comput Assist Radiol Surg 2025;20:981-9. [Crossref] [PubMed]
38. Takeda H, Akatsuka J, Kiriyama T, et al. Clinically Significant Prostate Cancer Prediction Using Multimodal Deep Learning with Prostate-Specific Antigen Restriction. Curr Oncol 2024;31:7180-9. [Crossref] [PubMed]
39. Chollet F. Xception: Deep Learning with Depthwise Separable Convolutions. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE; 2017:1800-7.
40. van Sloun RJG, Wildeboer RR, Mannaerts CK, et al. Deep Learning for Real-time, Automatic, and Scanner-adapted Prostate (Zone) Segmentation of Transrectal Ultrasound, for Example, Magnetic Resonance Imaging-transrectal Ultrasound Fusion Prostate Biopsy. Eur Urol Focus 2021;7:78-85. [Crossref] [PubMed]