**1. Introduction**

Prostate carcinoma (PCa) is the most common cancer and the third leading cause of cancer-related death among men in the United States [1]. A major challenge for PCa management is the lack of non-invasive tools that can differentiate aggressive versus non-aggressive cancer types [2]. This limitation can result in overdiagnosis and overtreatment, as evidenced by the fact that only one death is prevented for every 48 patients treated for PCa [3]. This overdiagnosis and overtreatment can lead to unnecessary biopsies, surgeries, radiotherapy, chemotherapy, and patient anxiety [2]. Better diagnostic methods could mitigate these unwarranted procedures. To meet this need for more effective screening, multiparametric magnetic resonance imaging (mpMRI) could be implemented to examine the entire prostate.

MpMRI of the prostate has been increasingly used for PCa screening in recent years [4]. MpMRI's current utility in screening stems from a high negative predictive value for prostate cancer. However, the full potential of mpMRI has not yet been achieved [5]. PCa overdiagnosis could be reduced with an mpMRI analysis that accomplishes better lesion detection, lesion classification (benign versus malignant), and lesion volume quantification.

Accurate prostate segmentation and volume estimation can provide invaluable information for the diagnosis and clinical management of benign prostatic hyperplasia (BPH) and PCa. This can improve BPH treatment, surgical planning, and predictions of PCa prognosis [6–8]. Prostate segmentation is necessary for magnetic resonance imaging (MRI) transrectal ultrasound (TRUS) fusion biopsy, which is increasingly used to diagnose PCa. MRI/TRUS fusion biopsy yield depends on accurate prostate segmentation on magnetic resonance images because the prostate edges form the reference frames for fusion with the ultrasound data [8]. Therefore, any inaccuracy in tracing prostate boundaries may lead to biopsy errors [9]. In addition to segmentation, prostate volume estimation is also a useful metric, especially with regard to BPH treatment, surgical planning, and PCa prognosis. BPH is one of the most common diseases that affects elderly men and reaches a prevalence of 90% by the ninth decade of life [10]. Large prostate volumes in men with BPH indicate a higher likelihood of more severe lower urinary tract symptoms and urinary retention [11–13]. Furthermore, studies have shown that patients have differential responses to BPH-targeted medications, depending on prostate size [11]. Additionally, prostate volume is considered when determining a surgical approach, with each procedure having its own risk profile [14]. In addition to guiding BPH treatment, prostate volume is used in PCa prognosis. Prostate size alone is a valuable marker for PCa prognosis; PCa is more accurately detected in prostates under 50 cm3 than in those over 50 cm<sup>3</sup> [6]. Prostate volume is also used to calculate prostate-specific antigen density, a figure that helps to differentiate BPH from PCa and can also be used to predict radical prostatectomy outcomes [15–17].

Accuracy of prostate lesion detection, segmentation, and volume estimation is important at different stages of PCa management. Lesion detection identifies regions for biopsy. Accurate segmentation is crucial for improved fusion biopsy yields. Additionally, segmentation improves radiotherapy delivery. Volume estimation predicts prognosis after prostatectomy [18–20]. Prostate lesion detection is crucial because the effective treatment of PCa directly depends on identifying cancer at its earliest stage [21–23]. Even though PCa most often follows an indolent course, it can show rapid progression in some cases. In these instances, lesion recognition on mpMRI is critical because it provides a region of high suspicion and a higher yield from targeted biopsy [24]. Without mpMRI lesion detection, random 12 core TRUS biopsies are performed, which may miss small or anteriorly located PCa [25].

While early prostate lesion detection improves timely PCa treatment, accurate lesion segmentation can improve radiotherapy [26]. Prostate lesion contouring is a major source of error when administering radiation therapy. This inexact segmentation can lead to the underdosage of the tumor as well as the overdosage of normal cells [20]. Although radiotherapy is an effective cancer treatment, its use is hampered by imprecise delineation. More precise contouring of a malignant lesion can improve lesion targeting and relative radiotherapeutic dosage, which can lead to lower recurrence rates [26,27].

Pre-operative prostate lesion volume estimation is a key metric for predicting the likelihood of positive surgical margins, biochemical prostate-specific antigen (PSA) recurrence, and cancer-specific survival post-prostatectomy [28–31]. This volume is a better indicator of surgical margins than other factors such as Gleason score and extracapsular extension [28]. Lesion volume also functions as an independent variable for PSA recurrence, an early sign of recurrent disease which may require salvage radiation therapy [32]. In addition, lesion volume predicts cancer-specific survival more accurately than variables such as lymphadenopathy, seminal vesical invasion, and Gleason score [29].

After prostate lesions have been detected on mpMRI, lesion characterization is important for selecting appropriate management options. Accurate prostate lesion classification on mpMRI could preclude biopsies in men with low-grade tumors, reduce the number of biopsy cores, and decrease the rate of overdiagnosis and false-negative biopsies [33]. Reduction in unnecessary biopsies is important, as potential TRUS biopsy complications include hematuria, lower urinary tract symptoms, and temporary erectile dysfunction [34]. Additionally, the number of biopsy cores obtained correlates with increased risk of complications, including rectal bleeding, hematospermia, bleeding complications, and acute urinary retention [34]. Furthermore, the overdetection of PCa exerts a major psychological toll on quality of life and increases the risk of overtreatment [2]. Overtreatment side effects that may occur after radical prostatectomy and radiotherapy include urinary incontinence, rectal bleeding and fistulae, and erectile dysfunction [2,35–37].

Artificial intelligence (AI) is a promising tool to improve prostate lesion detection, lesion characterization, and lesion volume quantification. AI can systematically evaluate mpMRI images [38]. Machine learning (ML), a branch of AI, and its sub-discipline, deep learning (DL), have become attractive techniques in medical imaging because of their ability to interpret large amounts of data [39]. By applying ML to prostate mpMRI data, imaging-based clinical decisions could be improved. The purpose of this review is to summarize ML applications for prostate mpMRI in regards to (1) prostate organ segmentation, (2) prostate lesion detection and segmentation, and (3) prostate lesion characterization.

#### **2. Multiparametric Magnetic Resonance Imaging**

Multiparametric magnetic resonance imaging (mpMRI) of the prostate is a form of advanced non-invasive imaging that combines standard anatomical sequences with functional imaging. It consists of T1-weighted images, T2-weighted images (T2W), and the following functional sequences: diffusion-weighted images (DWI) including the apparent diffusion coefficient maps (ADC) and dynamic contrast-enhanced images (DCE). Certain protocols also incorporate proton magnetic resonance spectroscopy imaging (MRSI) [40,41]. Typically, the functional techniques used are DWI and DCE. MRSI is more demanding than DWI and DCE because it requires more acquisition time, greater technical expertise, and intensive post-processing of the data. Therefore, it is not commonly used [42].

The advantages of seeing both the anatomy and functional ability of the prostate have made mpMRI an attractive imaging technique with many applications. It can accurately identify clinically relevant cancer. The combination of T2W, DWI, and DCE has high specificity, sensitivity, and negative predictive value in detecting PCa [43–45]. The use of all three functional sequences has been found to have a positive predictive value for PCa of 98% [46]. In addition to diagnosing PCa, mpMRI is also used in the management of the disease as the functional sequences aid in predicting tumor behavior. Prostate mpMRI has been used for active surveillance, tumor localization, staging, treatment planning, and monitoring of recurrence [40,41].

While mpMRI is a powerful imaging modality, it does have limitations. Differences in image acquisition techniques and protocols across institutions lead to heterogeneity in imaging quality and make it challenging to compare images [47]. Additionally, the learning curve for reading mpMRIs is steep, and there exists inter-observer variability [48–50]. The experience of radiologists reading these scans impacts the utility of prostate mpMRI images. In addition, the prostate gland is difficult to delineate, and various benign and pre-malignant processes can mimic PCa [51]. For example, the sensitivity for the detection of PCa in the transitional zone is limited by the heterogeneous nature of this zone in the setting of BPH, which can also exhibit increased cellularity further complicating the distinction. Furthermore, patient-related factors, including body habitus, prior procedures, and unconventional anatomy, can impact imaging. Artifacts, such as field inhomogeneity from rectal gas and metal implants, can substantially impede the interpretation and reporting of prostate mpMRI. Finally, it can be difficult to discriminate between post-treatment changes and local recurrence following treatment on mpMRI.

In an effort to assist in standardizing the acquisition, interpretation, and reporting of prostate mpMRI, the Prostate Imaging Reporting and Data System (PI-RADS) was developed by the European Society of Urogenital Radiology (ESUR) in 2012 [52]. The ESUR, in collaboration with the American College of Radiology and the AdMeTech Foundation, released updated versions PI-RADS v2 in 2015 and PI-RADS v2.1 in 2019 [53,54]. All of the PI-RADS versions offer guidance for protocols and specifications for image acquisition. The scoring systems provide frameworks to evaluate individual sequences of T2W, DWI, and DCE and to integrate these findings into an overall risk assessment category from 1 to 5. These risk categories assist in the determination of biopsies and the management of clinically significant PCa.
