Prostate cancer (PCa) is the second most common cancer among men worldwide, with approximately 1.4 million new cases and 375,000 deaths reported globally in 2020. While digital rectal examination and prostate-specific antigen (PSA) testing remain standard screening tools, multiparametric MRI (mpMRI) has replaced transrectal ultrasound (TRUS) as the first-line radiological screening modality due to its superior soft-tissue resolution and ability to provide multiple imaging parameters non-invasively.
The PI-RADS framework: The Prostate Imaging-Reporting and Data System (PI-RADS) provides standardized criteria for scanning, interpreting, and reporting mpMRI. PI-RADS scores range from 1 (low likelihood of clinically significant PCa) to 5 (high likelihood), and these scores guide biopsy decisions. Combining mpMRI with PI-RADS has significantly reduced overdiagnosis. However, manual interpretation of mpMRI remains complex, time-consuming, and subject to high inter-observer and intra-observer variability, meaning different radiologists may interpret the same MRI differently.
Enter deep learning: Computer-aided diagnosis (CAD) systems based on deep learning (DL) have emerged to address these interpretation challenges. CAD systems fall into two categories: computer-aided detection (CADe), which identifies and localizes possible lesions from the full mpMRI dataset, and computer-aided diagnosis (CADx), which evaluates tumor-suspected areas and assesses the aggressiveness of the cancer. This review surveys DL-based CAD applications across prostate cancer detection, segmentation, grading, radiotherapy planning, and prognostic assessment.
The review categorizes machine learning approaches into supervised learning (using labeled data), unsupervised learning (working with unlabeled datasets), semi-supervised learning (partially labeled), and reinforcement learning (feedback-based optimization). Deep learning, introduced by Hinton in 2006, is a subset of machine learning distinguished by its use of multilayer neural networks that automatically extract features from data, unlike traditional ML which relies on hand-crafted feature engineering by domain experts.
Key architectures: Convolutional neural networks (CNNs) are the dominant architecture for medical imaging. Their structure consists of convolutional layers (feature extraction via convolution kernels), max-pooling layers (dimensionality reduction), and fully connected layers (classification). Specialized CNN variants include AlexNet and ResNet for image classification, YOLO and Faster R-CNN for object detection, and U-Net and Mask R-CNN for semantic segmentation. Generative adversarial networks (GANs) are also prominent, consisting of a generator and discriminator network trained adversarially. Deep convolutional GANs (DCGANs) combine CNN and GAN strengths for improved CADx model design.
Evaluation metrics: Model performance is assessed using sensitivity (true positive detection rate), specificity (true negative identification rate), dice similarity coefficient (DSC, measuring segmentation overlap on a 0-1 scale), the Jaccard index, and area under the ROC curve (AUC). An AUC between 0.5 and 1 indicates predictive value, with values closer to 1 signifying superior performance. DSC values closer to 1 indicate better segmentation quality.
The typical CAD pipeline for prostate cancer involves imaging alignment, prostate localization and segmentation, feature-based lesion detection, and task-based classification. With growing volumes of labeled imaging data, supervised CNN models and unsupervised GAN frameworks have been incorporated into various stages of this pipeline.
Accurately distinguishing between low-risk and high-risk prostate cancer cases is critical to avoid overdiagnosis and delayed treatment. For low-risk patients under active surveillance, mpMRI serves as the primary imaging technique to monitor whether lesions have grown or metastasized. The review catalogs multiple DL-based detection models and their reported performance metrics.
Detection performance: Vente et al. developed a multitasking U-Net model using T2-weighted (T2W) and diffusion-weighted imaging (DWI) sequences capable of simultaneously detecting and grading PCa with excellent outcomes. Wang et al. designed an end-to-end CNN with two sub-networks (one for aligning apparent DWI and T2W, and another for classification), trained and assessed on 360 patients using fivefold cross-validation, achieving a sensitivity of 0.89 for identifying high-risk PCa. Ishioka et al. built a fully automated detection system using T2W data alone, demonstrating an AUC of 0.793.
DL vs. non-DL comparison: Wang et al. directly compared DL and non-DL algorithms for differentiating PCa using T2W sequences from 172 patients (79 with PCa, 93 with benign prostatic hyperplasia). The DL model achieved an ROC curve value of 0.84 versus 0.70 for the non-DL model. Sanford et al. trained a CNN on T2W/ADC/high-b values for PI-RADS scoring, confirming that DL models can match the diagnostic capability of clinical PI-RADS scoring by expert radiologists. Schelb et al. reported 93.0% accuracy and 0.95 AUC on a cohort of 250 + 62 patients using T2W, ADC, and DWI sequences.
Accurate segmentation of the prostate gland on MRI is essential for multiple clinical tasks: calculating prostate volume (PV) for PSA density measurements, planning nerve-sparing or fascial-sparing radical prostatectomy, delineating radiotherapy target areas, and monitoring tumor progression. However, segmentation remains challenging due to heterogeneity in MRI imaging quality and interference from adjacent structures like the bladder and rectum.
3D deep learning models: Zhu et al. developed a three-dimensional (3D) deep learning model with dense blocks to segment the prostate gland, achieving a DSC of 0.82 by exploiting relationships between adjacent MRI slices. Yan et al. proposed a backpropagation neural network integrating multi-level feature extraction, reaching a DSC of 0.84, which represented an average improvement of 3.19% over traditional random forest-based ML segmentation. To et al. used a 3D deep dense multipath CNN built from T2W and ADC sequences, achieving DSCs of 0.95 and 0.89 in two independent test sets.
Additional segmentation benchmarks: The review's Table 1 summarizes results from multiple groups. Ushinsky et al. achieved 0.898 DSC using T2W alone (287 patients). Bardis et al. reported 0.940 DSC on T2W data from 242 patients. Soerensen et al. obtained 0.92 +/- 0.02 DSC from 156 patients. Dai et al. developed a mask region-based CNN for both prostate gland and intraprostatic lesion segmentation, demonstrating the ability to automatically identify suspicious lesions directly from MRI without manual intervention.
Clinical significance: PSA density (PSA-D) is closely related to prostate volume and serves as an indicator of clinically significant PCa. Accurate DL-based segmentation enables more precise PV calculations than TRUS, which is susceptible to significant measurement errors when the prostate has an irregular shape. For surgical planning, precise boundary delineation is critical for preserving the neurovascular bundle during nerve-sparing prostatectomy and the pelvic fascia during fascial-sparing procedures.
Radiotherapy is a vital component of prostate cancer treatment, aiming to maximize the therapeutic gain ratio by delivering effective doses to the planned target volume (PTV) while minimizing exposure to organs at risk (OARs). However, anatomical variations caused by bladder and rectal filling, respiratory movements, and setup errors can displace PTV and OARs between treatment sessions. Image-guided adaptive radiotherapy (ART) has emerged to address these challenges.
ART modalities: The review describes three ART modes. Offline ART measures setup errors during initial treatments and adjusts subsequent fractions, but lacks flexibility for session-to-session anatomical changes. Online ART recalculates plans based on same-day imaging, improving accuracy but requiring more time. Real-time ART performs intra-fraction tracking and automatic plan adjustments during treatment, overcoming drawbacks of both offline and online approaches, though its safety requires further validation due to limited training databases.
Deep reinforcement learning for treatment planning: Sprouts et al. developed a virtual treatment planner (VTP) based on deep reinforcement learning using the Q-learning framework. Evaluated on 50 samples, it achieved a mean ProKnow plan score of 8.14 +/- 1.27, indicating its potential for intensity-modulated radiotherapy (IMRT) planning. Shen et al. introduced knowledge-guided DRL to enhance VTP training efficiency, improving the plan quality score to 8.82 +/- 0.29. Lempart et al. proposed a densely connected DL model based on a modified U-Net, trained on 160 patients for volumetric-modulated arc therapy dose prediction, maintaining mean percentage error within 1.9% for CTV and PTV and within 2.6% for OARs.
While MRI provides excellent soft-tissue contrast for tumor segmentation, it cannot directly supply the electron density maps or Hounsfield units needed for radiation dose calculation. Traditionally, this requires aligning MRI contours to CT images, a labor-intensive process prone to systematic errors that can diminish radiotherapy effectiveness. MRI-only radiotherapy solves this by converting MRI data into synthetic CT (sCT), enabling dose calculations directly from MRI.
Conventional sCT methods: Two commercial systems are in clinical use. MRCAT (Magnetic Resonance for Calculating Attenuation, by Philips) uses a bulk density assignment approach requiring multiple MRI sequences (air, liquid, and bone), with accuracy determined by segmentation quality. MriPlanner (by Spectronic Medical) uses an atlas-based approach requiring only a single MRI sequence, with accuracy determined by MRI-to-MRI alignment quality. MriPlanner has shown promising results in the MR-OPERA and MR-PROTECT clinical studies.
DL-based sCT generation: Deep learning models can generate highly precise sCT images in seconds. Fu et al. used 2D and 3D fully connected CNNs based on U-Net to generate pelvic sCT, completing the task in 5.5 seconds for 16 patients. Conditional GANs (adding a discriminator to U-Net) enhance sCT detail and robustness. CycleGAN (cGAN) architectures handle unpaired training data by incorporating additional generators and discriminators. Liu et al. proposed a multi-CycleGAN network with a novel Z-Net generator that achieved lower mean error and higher dose accuracy. The review's Table 3 shows sample sizes ranging from 16 to 40 patients and sCT generation times from 3.8 to 175 seconds depending on the model architecture.
Clinical impact: Combining MRI and CT simulations may reduce acute urogenital toxicity in PCa radiotherapy. DL-based sCT generation eliminates the need for manual CT-MRI alignment, reducing systematic errors and enabling faster, more accurate MRI-only treatment workflows.
Monitoring patients after radical prostatectomy and during active surveillance is critical for improving prostate cancer outcomes. PSA remains the primary biomarker for biochemical recurrence (BCR), as recognized by the European Association of Urology guidelines. However, PSA levels can fluctuate and be influenced by various factors, creating diagnostic bias. MRI-based assessment of recurrence offers non-invasive anatomical insights that complement PSA monitoring.
BCR prediction with radiomics: Yan et al. conducted a multicenter study using a novel "deep radiomic signature" model that combined quantitative features and radiomics extracted from prostate MRI with DL-based survival analysis. Evaluated on approximately 600 patients who underwent radical prostatectomy, the model achieved maximum AUC values of 0.85 for BCR-free survival prediction at 3 years and 0.88 at 5 years, demonstrating strong predictive performance across multiple centers.
Bone metastasis detection: Over 80% of patients with advanced prostate cancer develop bone metastases, and conventional bone scintigraphy has questioned accuracy and sensitivity. PSMA PET-CT and MRI show potential for earlier metastasis detection. Liu et al. detected and segmented pelvic bone metastases using dual 3D U-Net DL algorithms relying on T1-weighted and diffusion-weighted imaging sequences. Through two evaluation rounds, they achieved a mean DSC above 0.85 for pelvic bone segmentation and a maximum AUC of 0.85 for metastasis detection.
Broader applications: DL has also been applied to Gleason score prediction from pathological sections with diagnostic power equal to that of pathologists. DL-based auto-contouring algorithms for radiotherapy reduce workload and inter-observer variability across multiple radiotherapy centers. Integration of DL with nomograms now enables inclusion of variables such as PSA, prostate volume, patient age, free/total PSA ratio, and PSA density into MRI-based diagnostic workflows.
2D vs. 3D processing: Most current DL algorithms rely on 2D images for feature extraction and analysis, meaning they may not adequately capture 3D spatial anatomical information from clinical imaging. While DL has been used for 3D segmentation in liver and cardiovascular imaging, research on 3D prostate gland segmentation remains scarce. Developing computational methods that work effectively with 3D medical images while maintaining high detection performance is a significant technical challenge.
Limited multimodal and multisequence data: Most prostate MRI studies include only T2W and DWI sequences. Despite the diminishing role of dynamic contrast-enhanced (DCE) imaging according to recent PI-RADS guidelines, incorporating ADC maps and fusing multiple modalities for 3D tumor segmentation could further enhance CNN accuracy. The effectiveness of DCE remains debated due to its time-consuming nature and the risk of nephrogenic systemic fibrosis. The review recommends prioritizing biparametric MRI (T2W + DWI) for rapid screening.
Explainability and interpretability: A primary limitation of DL across all fields is the incomprehensibility of predictions and decisions. In medical contexts where DL-based decisions carry life-threatening consequences, the rationale behind conclusions must be transparent. The authors call for developing "explainable AI" (XAI) that provides accurate predictions with understandable assessment criteria. Beyond explainability, understanding complex biological contexts (molecular mechanisms, genetic expression, cellular microenvironments) is essential for novel biomarker development and disease pathogenesis research.
Digital biopsy: The review introduces the concept of "digital biopsy," which involves analyzing digital images and identifying characteristic features focused on tumor heterogeneity rather than contour, using multi-omics computational power. The authors propose that DL-based digital biopsy could become the "next-generation biopsy" for low-risk PCa patients, offering non-invasive disease assessment and prediction. Additional needs include expanding annotated medical databases, optimizing model architectures through parallelized sub-networks, and combining multi-omics data for comprehensive morphological analysis.