A review on innovations in prostate cancer diagnosis: automated techniques for gleason score estimation via mpMRI and DWSI imaging

Recent advances in DL have significantly improved prostate cancer detection using mpMRI, focusing on segmentation, classification, and risk assessment. Machireddy et al. [16] developed a U-Net with attention, achieving improved malignancy detection across 55 biopsy-confirmed cases. However, the model lacked integration of ADC and T2W modalities, limiting its diagnostic depth. Similarly, Soni et al. [17] offered SEMRCNN that exploited both ADC and T2W images through parallel convolutions. While enhancing segmentation, it suffered from poor lesion boundary precision due to overlapping tissue. In both cases, lesion detection was improved but not fully reliable across tissue types. The absence of multimodal harmonization reduced the models’ real-world robustness. Therefore, combining features across image sequences remains an open avenue.

In a related effort, Jiang et al. [18] introduced MiniSegCaps, a lightweight capsule-based architecture for segmentation and PI-RADS classification. The model effectively distinguished between high- and low-grade lesions using fewer parameters. However, dataset limitations prevented it from executing detailed lesion grading or stratification. Similarly, Wang et al. [19] proposed SegDGAN, a GAN-driven segmentation model that delivered high Dice scores. Nonetheless, false positives were frequent due to poor delineation of tumor margins, confusing healthy tissue for cancerous regions. Both approaches reflect how lightweight and adversarial models can enhance detection. Yet, segmentation fidelity remains hindered by anatomical complexity and limited datasets. This highlights the need for models that pair accuracy with boundary sensitivity across clinical variations.

Fusion-based models have also been explored to improve visibility and diagnostic consistency in mpMRI. Huang et al. [20] designed a U-Net-based image fusion method to enhance lesion contrast by combining multiple MRI sequences. Its success, however, was heavily influenced by image quality, making it less effective under noisy or variable scan conditions. Likewise, Gavade et al. [21] applied U-Net for ROI segmentation, followed by classification, stressing the importance of interpretability. These models indicate that fusion improves feature representation but depends greatly on acquisition consistency. They also call attention to the need for standardized imaging pipelines. While accuracy was enhanced, clinical applicability was constrained. Thus, deep fusion remains promising but requires robustness across scanners and centers.

Hung et al. [22] developed a Cross-slice Attention Transformer (CAT) that addressed anatomical precision by attending across 2D slices of T2W MRI scans. The model improved segmentation of transitional and peripheral zones but was limited by its reliance on T2W alone. In parallel, Mehmood et al. [23] adopted EfficientNet-based transfer learning across T2W, ADC, and DWI using a triple-branch network. This multi-sequence integration improved classification stability under diverse imaging conditions. Both models exemplify the value of sequence-aware transformers and pretraining. Yet, challenges remain in generalizing to unseen institutions or varied equipment. Incorporating attention has shown meaningful gains, but external validation is often absent. More studies must explore how transformer backbones perform in real diagnostic environments.

Yi et al. [24] proposed a deep neural network (DNN)-based computer-aided diagnosis (CAD) system that incorporated multiple mpMRI modalities. While showing potential, the method lacked advanced scoring features for atypical presentations, reducing sensitivity in less common cases. Similarly, Duran et al. [25] developed ProstAttention-Net, a CNN architecture that fused lesion grading with segmentation using attention mechanisms. Its end-to-end design was effective, yet it only utilized mpMRI, missing potential enhancement from hybrid imaging like PET/MRI. Both approaches reflect the growing trend of joint learning frameworks in CAD. However, their clinical reliability still hinges on modality expansion and performance across patient subgroups. These models reveal that multi-task learning is feasible but requires stronger validation.

Table 1 depicts the techniques, advantages, and limitations in the context of prostate cancer detection and segmentation using mpMRI. In this study, several DL techniques have been applied, such as U-net with attention mechanisms, SEMRCNN, and SegDGAN, for prostate cancer detection and segmentation from mpMRI images. The models provide advantages in terms of improved accuracy, better lesion localization, and enhanced segmentation through the integration of multi-modal data along with attention mechanisms. Despite all of the above, there are some challenges: the segmentation cannot be done properly because the boundary overlaps with the normal tissue; the reliance on very few imaging modalities like T2WI; and false positives. Although some methods like ProstAttention-Net and MiniSegCaps improve the performance of segmentation and classification, they still have problems regarding data quality and require further multi-modal integration. The problems here include inconsistent data, small sample size, and misclassification due to incomplete feature extraction. Image quality remains problematic, and cancer boundaries are far from clear; other modes of imaging are usually required to improve detection and segmentation.

Zhong et al. [26] presented a convolutional recurrent neural network for diffusion-weighted imaging (CRNN-DWI) that combines CNNs for spatial feature extraction and RNNs for temporal or directional analysis to reconstruct heavily undersampled DWI datasets. While effective in enhancing reconstruction quality, the model does not directly support prostate cancer segmentation. Integration with segmentation frameworks is required to provide a complete diagnostic solution. This limitation reduces its standalone utility in end-to-end clinical workflows. Similarly, Hu et al. [27] compared two DWI acquisition strategies zoomed-field-of-view (zDWI) and full-field-of-view (fDWI), within DL-based CAD systems. Although improved delineation was achieved using zDWI, the model’s performance lacked validation across diverse patient populations. This indicated poor generalizability, restricting broader application. Further research was encouraged to strengthen model resilience for robust real-world deployment.

Kaye et al. [28] explored the acceleration of prostate DWI imaging by reducing averages and applying a guided DnCNN denoising network to preserve image fidelity. The absence of a true noise-free reference for DWI or ADC maps necessitated the use of surrogate high b-value images, which introduced training uncertainty. This limitation led to inconsistent results, especially for subtle lesions. In response, the authors recommended exploring noise2noise training paradigms that do not require clean ground truths. Similarly, Abdelmaksoud et al. [29] introduced a CAD system using non-negative matrix factorization (NMF) for integrating multi-b-value DWI-derived features. Although their method accurately localized the prostate region, statistical analysis for optimal b-value selection was not conducted. Reducing b-values could improve computational efficiency without significant accuracy loss. These studies underscore the importance of balancing image quality with acquisition speed and processing demands.

In related work, Mao et al. [30] implemented a 3D U-shaped CNN model for automatic prostate segmentation using DWSI. The approach used multi-b-value DWSI images to capture rich diffusion contrast, aiding in robust tissue delineation. Furthermore, uncertainty maps were employed to highlight regions of confident predictions, improving treatment planning reliability. However, the model did not include derivative maps like ADC or b0 gradients, limiting its diagnostic completeness. Likewise, Liu et al. [31] developed a 3D U-Net CNN for lymph node (LN) detection from pelvic DWSI images in prostate cancer cases. While the model successfully segmented suspicious LNs >0.8 cm, its use was limited to the pelvic region. Additional validation was recommended to generalize the method to other anatomical regions and patient cohorts. These studies highlight the emerging role of 3D CNNs in diffusion-weighted segmentation but reveal application scope limitations.

Alhassan [32] proposed a novel system that segments prostate lesions using preprocessing and ROI-based extraction to isolate potentially malignant areas. The system was built upon a multi-level bidirectional long short-term memory network (Bi-LSTM) with attention mechanisms to refine focus within suspicious regions. Although promising in theory, the method incurs high computational overheads, making it unsuitable for real-time deployment without powerful hardware. In continuation, Jeong et al. [33] applied deep learning reconstruction (DLR) to DWI enhancement, improving lesion visibility and signal-to-noise ratio (SNR). Their study also noted the potential for reducing scan duration without degrading diagnostic accuracy. However, the clinical impact of DLR, including long-term patient outcomes or treatment modification, remains unexplored. Therefore, while reconstruction methods improve visual quality, their contribution to diagnostic or prognostic accuracy needs further clinical correlation.

Hu et al. [34] presented a GAN-based framework for synthesizing high-b-value DWSI images from standard acquisitions to enhance lesion conspicuity. The system offered a pathway to improved diagnosis without the need for extended scan times. However, performance under moderate to severe motion artifacts was not evaluated, limiting its reliability under real-world scanning variability. Similarly, Motamed et al. [35] proposed a transfer learning-enhanced U-Net for segmenting prostate zones, namely, the whole gland (WG) and transitional zone (TZ). While segmentation accuracy was achieved, the study did not provide prostate volume calculations—a crucial clinical parameter for prognosis and therapy monitoring. Together, these efforts demonstrate that while advanced DL architectures can improve imaging outputs, practical clinical metrics and generalizability are essential for full adoption.

Clearly from Table 2, it has been noticed that some techniques of DL have used DWSI in prostate cancer diagnostics such as CRNN, 3D U-Net CNN, and Modified U-Net. Among the quality enhancement in terms of detection improvement of cancer through their increased segmentation accuracy of the cancer area, important gains occurred over accelerated scan time which were needed for use in a clinic. However, disadvantages include issues like the problem of having a noise-free ground truth for training, inadequate validation across various patient populations, and necessity for further model refinement when applying it clinically. The two approaches with integration of uncertainty quantification and multi-modal learning approaches present promising outcomes but are mostly computationally expensive, limiting the practicability in real-time. Further research and improvements are needed to overcome these challenges and enhance clinical applicability for prostate cancer detection.

Chaddad et al. [36] proposed a radiomics approach to estimate GS using deep entropy features (DEFs) derived from mpMRI through nine pretrained CNNs, such as NASNet-Mobile. These features captured tissue texture and variability from intermediate layers of CNNs, enhancing lesion characterization. However, the method fell short in capturing complex non-linear feature interactions. This limitation reduced its predictive robustness in high-grade tumor cases. Similarly, Butt et al. [37] developed a multi-label ensemble deep-learning classifier for GS prediction in histopathology using the SICAPv2 dataset. Although patch-level accuracy was enhanced, pixel-level granularity was absent, compromising diagnostic precision. Moreover, majority voting-induced label inconsistencies remained an issue. Future extensions may address pixel-wise classification for finer GS discrimination.

Pellicer-Valero et al. [38] introduced a fully automated framework using 3D Retina U-Net for mpMRI-based prostate cancer detection, segmentation, and Gleason Grade Group (GGG) estimation. The RetinaNet component facilitated one-shot lesion localization, while the U-Net managed pixel-wise delineation. Classification and bounding box refinement were guided by intermediate activation maps. While this multi-task model demonstrated promising GGG prediction, integration with clinical workflows remains to be validated. In continuation, Yoshimura et al. [39] combined semantic segmentation with a 3D-CNN to estimate GS directly from diagnostic MRI. The method aimed to replace invasive biopsies with non-invasive, image-driven assessment. While it improved accuracy through tissue segmentation, its broader application in early diagnosis and treatment guidance is still evolving.

Khalek et al. [40] emphasized the strong correlation between DWI/ADC parameters and GS for prostate cancer aggressiveness evaluation. Their findings underscored the potential of quantitative diffusion metrics as imaging biomarkers for tumor grading. However, clinical applicability demands further study to improve reproducibility and standardization. Similarly, Li et al. [41] developed a CNN-based architecture that used multiscale standard convolutions and atrous spatial pyramid pooling (ASPP) to segment and grade Gleason patterns. Post-processing via conditional random fields (CRFs) enhanced precision, but also increased computational burden. Although segmentation performance improved, balancing efficiency and clinical scalability remains a challenge. These works reinforce the growing role of hybrid CNN models in pathology-grade prediction.

Mary et al. [42] proposed a hyperparameter-optimized deep belief network (CDBN-EHO) for joint Gleason grading and prostate cancer classification in histopathological images. Their approach combined Elephant Herding Optimization with a multi-task prediction framework, improving training efficiency. Nonetheless, the method showed weaknesses in subtle or heterogeneous patterns, often misclassifying poorly differentiated tissues. Similarly, Nai et al. [43] integrated mpMRI and PET features with ML classifiers like SVM, kNN, and Ensemble Models for GS prediction. Despite achieving high accuracy using eight selected radiomic features, essential discriminative traits may have been omitted. The robustness of this minimal feature set across diverse clinical scenarios warrants further scrutiny.

In related work, Balaha et al. [44] introduced a two-stage DL pipeline for prostate cancer classification and segmentation using mpMRI. The framework first localized diseased areas via U-Net-based segmentation, followed by classification for accurate diagnosis. This combination enabled early identification and delineation of malignant zones, supporting targeted therapy planning. However, generalizability to unseen datasets and integration with biopsy data were not addressed. Furthermore, real-time inference speed and hardware requirements remain critical for clinical application.

As illustrated in Table 3, various techniques have been devised for the estimation of GS in prostate cancer by using DWSI and mpMRI imaging. These techniques include DEFs with CNNs, multi-label ensemble DL classifiers, and 3D Retina U-Net, which contribute to significant improvements in texture analysis, label consistency, and fully automated segmentation in the estimation of GGG. However, these techniques have limitations such as a failure to capture non-linear correlations, loss of finer pixel-level details, struggling with complex or noisy images, and increased computational complexity. Although semantic segmentation and CNN-based region segmentation improve the grading accuracy, there is still room for further refinement. DWI/ADC parameters indicate promising aspects of assessing cancer aggressiveness, but it needs more standardization. ML models involving radiomic features reduce inter-reader variability but can err in some clinical situations. Further work is still needed in order to sharpen and expedite these techniques for estimating the GS.

Varan et al. [45] introduced a binary classification framework for prostate cancer detection by integrating ridge regression-based feature selection with a linear SVM. Four selection strategies correlation coefficient, forward-sequential, backward-sequential, and significance-based ensured robustness and relevance in chosen features. Their method improved training performance and generalization. However, reliance on linear SVM restricted the capture of non-linear data relationships. Extending the approach with kernelized models may enhance performance. Similarly, Akinnuwesi et al. [46] developed SVM-PCa-EDD for early differential diagnosis from non-imaging clinical datasets. Despite promising classification of positive and negative patients, harmonizing multiple feature selection outputs proved complex and time-consuming.

Fan et al. [47] devised a radiomic approach combining mpMRI and clinical data to predict biologically relevant features of prostate cancer, including Ki67, S100, and perineural invasion (PNI). After normalization, Reconstruction Feature Elimination with RFs was used for selecting key features. While insightful, manual lesion segmentation introduced interobserver variability, limiting reproducibility. Gaudiano et al. [48] investigated ML-enhanced radiomics to improve PI-RADS-3 lesion classification using ADC map features. Feature selection used least absolute shrinkage and selection operator (LASSO) and Wilcoxon tests before SVM classification. However, failure to stratify peripheral and TZs reduced specificity. Anatomical variation in lesion characteristics warrants zone-wise modeling in future work.

Jin et al. [49] presented a voxel-wise prostate cancer classifier using a Bayesian model that incorporated spatial correlation and patient-level heterogeneity. By integrating Gaussian processes, autoregressive models, and hierarchical priors, this method improved lesion localization. Although effective, the model operated on 2D slices due to limitations in co-registering 3D MRI with histopathology. In related work, Chen et al. [50] developed ML-based radiomics models using ADC, DWI, and T2-weighted data to predict clinically significant prostate cancer (csPCa) and general prostate cancer. RF and multilayer perceptron models were used. Manual segmentation increased variability and labor overhead, necessitating automation for scalability.

Tang et al. [51] developed a computer-aided diagnostic system utilizing mpMRI-derived geometric, topological, and texture features. The extracted features Minkowski functionals, gray-level co-occurrence matrix/gray-level gradient co-occurrence matrix (GLCM/GLGCM) textures, and wavelet descriptors were input into ML classifiers like KNN and SVM. While effective, generalization to diverse patient populations remains a challenge due to specificity of extracted features and model dependencies. In a similar domain, Zandie et al. [52] investigated mpMRI-based radiomic models to predict GGGs using optimal classifier-feature combinations. Despite reduced biopsy needs, absence of augmentation strategies limited class balance and overall model robustness.

Garg and Juneja [53] proposed using Possibilistic Exponential Spatial Fuzzy Clustering (PESFC) for malignant prostate capsule segmentation based on ADC, dynamic contrast-enhanced (DCE), and T2w imaging. The model leveraged spatial and fuzzy information to detect early-stage malignancies, contributing to reduced mortality. However, accurately distinguishing blurred boundaries between malignant and benign tissues remained a core limitation. Addressing this could further enhance diagnostic precision. Overall, these studies reveal strong advances in ML and radiomics for prostate cancer detection, while highlighting the persistent challenges in segmentation accuracy, model generalizability, and clinical integration across diverse data modalities.

As mentioned in Table 4, various ML techniques have been used to provide improved diagnosis of prostate cancer using mpMRI and DWSI imaging. Such methods include ridge regression with linear SVM, SVM-PCa-EDD, and Bayesian spatial models in enhancing diagnostic potential with regard to feature relevance, early diagnosis, and cancer localization. Radiomics, RF, and recursive feature elimination (RFE) allow for the non-invasive prediction of traits of the cancer, but manual segmentation continues to suffer from heterogeneity. Techniques like PESFC have improved segmentation accuracy, yet challenges remain with defining cancer boundaries. Overall, while these approaches show promise, limitations they face include a need for manual segmentation, difficulties in generalizing, and sub-optimal representation of non-linear complex relations among the features. Added to this is difficulty in handling high-dimensional 3-D MRI data, thereby keeping malignancy borders accurate. Such challenges make these ML-based approaches very much the basis for further fine-tuning and optimization.

Sethi et al. [54] proposed an LSTM-DBN-based model for evaluating gene expression in prostate cancer diagnosis and classification. The model learned complex interactions in gene-expressed data and was optimized using the Enhanced Whale Herd Optimization (EWHO) method. This ensured high accuracy and sensitivity for molecular-level detection. In continuation, Xu et al. [55] developed a fully automated CNN-based system using U-Net for detecting metastatic prostate lesions in PET/CT images. A weighted Dice loss function improved segmentation of lesions across diverse anatomical contexts. However, lesions near the bladder posed challenges due to signal overlap. Addressing this proximity issue is crucial for improving detection precision and clinical reliability.

Liu et al. [56] combined S-Mask R-CNN and Inception-v3 to detect prostate cancer from ultrasound images with improved segmentation. The approach enabled accurate identification of candidate regions, addressing limitations in 2D-to-3D ultrasound transitions. However, handling volumetric data remains computationally intensive and may hinder real-time application. Similarly, Rovera et al. [57] used k-means clustering for lymph-node identification in intraoperative 68Ga-PSMA-11 PET/CT scans. Although effective in segmenting nodal structures, the approach lacked the ability to detect extra-nodal disease in fatty tissues. Incorporating pathological variability could enhance this model’s diagnostic comprehensiveness.

Hassan et al. [58] developed a multi-modal DL classifier integrating pretrained DL models for prostate cancer detection using ultrasound and MRI data. This architecture improved performance across different image modalities. However, inconsistencies in imaging protocols and equipment across clinical setups reduced model generalizability. In related work, Jiang et al. [59] introduced MicroSegNet for high-resolution prostate segmentation in micro-ultrasound images. While it improved segmentation precision, its dependence on specific hardware limited its adaptability. Broader applicability would require fine-tuning for a range of devices and imaging configurations in clinical use.

Singh et al. [60] utilized a 3D-CNN model to detect prostate cancer and estimate GS from MRI data, focusing on epithelial tissues. The model localized lesions with high accuracy but struggled with small or anatomically complex tumors. Missed detections near difficult-to-image zones such as the bladder remain a challenge. Similarly, Yuan et al. [61] proposed Z-SSMNet, a zonal-aware self-supervised mesh network using bi-parametric MRI. It captured both inter- and intra-slice variations through 2D, 2.5D, and 3D CNN modules. However, reliance on zonal masks from specific T2-weighted imaging/apparent diffusion coefficient (T2WI/ADC) configurations reduced generalizability when data properties changed significantly.

Abdelmaksoud et al. [62] introduced a CAD framework that detected prostate cancer using DWI images collected at multiple b-values. Feature extraction was performed using NMF to isolate patterns indicative of cancerous tissue. While segmentation accuracy improved, the system’s dependency on DWI image quality limited robustness. Similarly, Korevaar et al. [63] suggested a 3D-CNN model for incidental detection of prostate cancer in routine abdominal CT scans. The framework leveraged spatial patterns in volumetric data for cancer identification. Despite its strong performance, low soft tissue contrast in CT images remains a limiting factor for detection accuracy.

From Table 5, it becomes clear that various imaging modalities are employed for the diagnosis of prostate cancer, with each modality having its advantages and drawbacks. An analysis of gene expression data provides an advantage in modeling sequential dependencies and hierarchical characteristics but is limited to non-imaging data. PET/CT enables better visibility of lesions, yet the proximity of lesions to the bladder complicates interpretation. The ultrasound and micro-ultrasound modalities demonstrate high-definition segmentation; yet, there are constraints posed by 3D data and device configuration. It gives a good quality of high-contrast images of the cancer with MRI and DWI; however, some small cancers are still missed or masked by artifacts. Again, the CT scan, although very useful, will not provide the same soft tissue contrast. Combining these modalities with DL algorithms shows better prospects for enhancing the accuracy of diagnosis.

The success of DL and ML methods in prostate cancer diagnosis from imaging data is often demonstrated through performance metrics such as accuracy, specificity, and Dice similarity coefficient. This section highlights a methodological evaluation based on the use of real clinical datasets, the level of external testing, and whether they were compared against expert or biopsy-confirmed outcomes.