Intelligent Computer-Aided Prostate Cancer Diagnosis Systems: State-of-the-Art and Future Directions

Prostate Cancer (PCa) is one of the common cancers amongmen in the world. About 16.67% ofmenwill be affected by PCa in their life. Due to the integration of magnetic resonance imaging in the current clinical procedure for detecting prostate cancer and the apparent success of imaging techniques in the estimation of PCa volume in the gland, we provide amore detailed review ofmethodologies that use specific parameters for prostate tissue representation. After collecting over 200 researches on image-based systems for diagnosing prostate cancer, in this paper, we provide a detailed review of existing computer-aided diagnosis (CAD) methods and approaches to identify prostate cancer from images generated using Near-Infrared (NIR), Mid-Infrared (MIR), and Magnetic Resonance Imaging (MRI) techniques. Furthermore, we introduce two research methodologies to build intelligent CAD systems. &e first methodology applies a fuzzy integral method to maintain the diversity and capacity of different classifiers aggregation to detect PCa tumor from NIR andMIR images.&e secondmethodology investigates a typical workflow for developing an automated prostate cancer diagnosis using MRI images. Essentially, CAD development remains a helpful tool of radiology for diagnosing prostate cancer disease. Nonetheless, a complete implementation of effective and intelligent methods is still required for the PCa-diagnostic system. While some CAD applications work well, some limitations need to be solved for automated clinical PCa diagnostic. It is anticipated that more advances should be made in computational image analysis and computer-assisted approaches to satisfy clinical needs shortly in the coming years.


Introduction
Cancer is one of the most critical health issues globally, in terms of morbidity, mortality, and its social, economic, or quality of life, affecting one in three people throughout their lives [1]. For all age groups, cancer became the leading cause of death worldwide for men and women after cardiovascular diseases [2]. Prostate cancer is the most frequent tumor location in men (excluding nonmelanoma skin tumors) and the third leading cause of death from cancer, in both cases behind cancer of the lung and colorectal. It is estimated that one in six men will develop prostate cancer in their lifetime [3]. e probability of developing prostate cancer increases with age so that nine out of ten cases appear in people over 65 years of age [4].
Diagnostic practices and therapeutic options have continued to evolve for detecting prostate cancer. Precisely, recent advances in prostate imaging make it possible to detect tiny tumors and guide treatment [5]. Computer-Aided Diagnosis (CAD) software and tools are designed to aid physicians in diagnosing suspicious areas of the image. e principle is to extract from one or more types of medical images (such as MRI, CT, US, and PET) discriminating characteristics of the pathology than to develop a prediction model from a database learning. is empirical model then makes it possible to quantify the probability that an area of interest or a voxel of a test image is pathological [6].
is work's main contributions and goals are to conduct a detailed review of existing CAD methods and approaches for identifying prostate cancer from images. Additionally, we proposed two research methodologies of CAD systems based on Near-Infrared (NIR), Mid-Infrared (MIR), and Magnetic Resonance Imaging (MRI) techniques. e rest of the paper is organized as follows. Section 2 provides detailed background on prostate cancer. Section 3 describes the near-infrared (NIR) and mid-infrared (MIR) images. Section 4 describes the two research methodologies of a CAD system on NIR, MIR, and MRI images. Section 5 gives a detailed explanation of the clinical application. In Section 6, we discuss our observations and list the future directions. Finally, we conclude our work in Section 7.

Background
e prostate is one of the essential glands in humans that is known as a walnut-sized gland [1]. e anatomical structure of the prostate gland is divided into three standard zones. e peripheral zone (PZ) is the most significant part of the gland that accounts for 70% of a gland volume and extends from the base to the apex of the prostate gland. e second part is the central zone (CZ), representing 25% of the gland, and the transition zone (TZ) represents the 5% rest of the prostate. e cancer of the prostate is commonly diagnosed at the PZ, where around 70% arises. On the other hand, 10% to 20% of prostate cancer cases are affected by a tumor in the TZ. About 5-10% prostate cancer is located in CZ [2].
Since the late 1980s and 1990s, there is a revolutionary shift in the health care agenda and diagnosis techniques of the developed world. Cancer starts taking place in the leading causes of death in humans. In the report [3], cancer named to be the second cause of death after heart diseases. Within different forms of cancer, the most prevalent source of male cancer and the second leading cause of cancer death after lung cancer is Prostate Cancer (PCa). Among men, prostate cancer is a globally common cancer. It grows rapidly [4]. erefore, prostate cancer and its diagnosis techniques make a priority to prostate cancer research. According to the American Cancer Society Report of 2019, prostate cancer cases in the US up to 174,650 patients will be diagnosed, and the estimated number of deaths will reach 31,620 cases. Current studies have predicted that the risks of cancerous disease in Asia will be rising in the future, especially in the growth countries [5]. e paper [6] anticipated that cancer would significantly increase on the continent of Asia, where 60% of the global citizens live. erefore, macro safety strategies in the future are a significant need for dealing with cancer. Early diagnosis of prostate cancer plays an essential role in survival rate or chronic conditions. e case of a population screening process has been a topic of continuous discussion since the early 1990s [7]. PCa is one of the leading cancers causing death through the world [8]. Due to the limitations in the existing clinical practices, accurate and easy diagnosis and assessment of PCa still needed improvements. Computer-aided diagnostic approaches would probably play an influential role in addressing this problem in the near future [9]. Nonetheless, since the early nineties, the need for community screening had been the focus of ongoing debate [7]. In medical tests for prostate cancer, the immunoassay-based assessment of blood levels of Prostate Specific Antigen (PSA) is generally accepted as the first phase in clinical screening for prostate cancer. Virtual (DRE) and transrectal ultrasounds (TRUS) are usually taken for patients with a higher PSA level.
Many large-scale prostate types of research focused on specific antigen studies, and they have shown reduced death risk related to PCa [10][11][12]. Unfortunately, as these approaches are simplistic and lack precision, assessing the occurrence and diagnosis of the disease is now carried out exclusively through a histopathological review of biopsy samples collected through TRUS, or MRI guidelines. However, some major studies have reported contradictory results, suggesting that PSA screening has no critical impact on PCa death rates [13,14]. Due to the less importance of the PSA in the diagnosis process, it has been issued or marked with the "D" letter symbol by the United States Preventive Services Task Force (USPSTF). It culminated in a shift away from the use in the United States [15]. In detecting clinically significant tumors, the random biopsy approach is also vulnerable to low sensitivity [16][17][18][19]. Although TRUS is more reliable and cost-effective than mp-MR, its low sensitivity does not render a broad patient population [20].
rough this restriction, the imaging of endorectal MR leads to the PSA, DRE, and biopsy outcomes despite localization and measurement of cancer scale and extension [21,22].
For endorectal images, the ROC curve for locating cancer is higher for apex, center, and base than for DRE. Additionally, endorectal MR imaging is more reliable in ingrowth and prostate early diagnosis, especially for the prostate base against TRUS controlled biopsy in the localization of the tumor [23]. is paper emphasis the focus of Magnetic Image Resonance (MR) Images more than transrectal ultrasound (TRUS) because of the success of MR in estimating the volume of PCa. e use of TRUS gives underestimate calculations. erefore, the volume is used to calculate PSA density, and TRUS leads to inaccurate analysis. MR provides a more accurate estimation of prostate volume compared to TRUS [24]. Identifying the exact location of prostate cancer using transrectal ultrasound is inaccurate as more as MRI images. e recent prospective study presented that the sensitivity of the diagnosis process of a model that uses transrectal Ultrasound (TRUS) to diagnose PCa is only 70.4% [25].
Song et al. [26] used deep convolution neural networks to perform a classification task on 195 patients with localized PCa collected from a prostate database. e data were sequence of T2W, DWI, and ADC combinations. It was related to two classes of cancerous data divided to 215 for training and 23 for validation, and 23 for testing; the second class was healthy people, and it is divided into 229 training and 25 for validation, and 32 for testing. After implementing DCNNs, the outcome accuracy reached 94%. Treviño et al. [27] examined rapid perceptual processing on T2W images, and the rapid radiological perception could be observed using T2W multislice imaging. e MR has the ability to extract the perceptual processing quickly and could also detect and localize lesions even in complex imaging modalities.
Chirra et al. [28] used about 147 patient T2-weighted MR datasets downloaded from four different sites containing 406 three-dimensional voxel-wise radiomic features from five different families (gray, Haralick, gradient, Laws, and Gabor) and evaluated in a reproducible either tumor or nontumor and the discrimination process of tumor regions, and they reached an accuracy of ≈0.8 in the classification phase. A recent review [29] concluded the independent behavior of the sensitivity in the diagnosis process of PCa on MR mages and many factors such as the enrollment of patient (consecutive or not consecutive), the strength of MR images (1.5 Tesla or 3 Tesla), the reference standard of bias risk (low risk of bias and high risk of bias), and whether readers are blinded or not to the histological findings (blind or not blinded).

Near-Infrared (NIR) and Mid-Infrared (MIR) Images.
With the assistance of Maestro in vivo imaging system, imaging is conducted. e Near-Infrared (NIR) and Mid-Infrared (MIR) images are acquired by using the appropriate filter set (deep red filter set for Prostate Specific Membrane Antigen (PSMA)-1-IR800 and yellow filter set for PSMA-1-Cy5.5). During this process, different points are chosen for imaging. e temperature of 37°C is tuned for the imaging bed through imaging operation. A nose cone is adjusted with an imaging bed for isoflurane inhalation.
Fluorescent molecular tomographic (FMT) images are obtained by using the FMT2500 device (Perkin-Elmer, Waltham, MA), and the three-dimensional reconstructions of fluorescent signals are captured using the accompanying software, TrueQuant. Quantification of fluorescent signals is also obtained by calibration of PSMA-1-IR800 and PSMA-1-Cy5.5 by the 780 nm and 680 nm channels, separately.
Several similar NIR/MIR images are collected by developing PSMA-targeted NIR/MIR optical imaging probes. ese are used for analyzing and visualization of prostate cancer intraoperatively. A high affinity PSMA ligand (PSMA-1) is synthesized with low molecular weight and further labeled using commercially available NIR/MIR dyes: Cy5.5.4 and IRDy800 demonstrated the utility of such probes to selectively bind of the prostate tumor in targeting both orthotropic and heterotopic prostate tumors.

NIR and MIR Acquisition.
Intuitively, image acquisition is the first stage of any vision system. In the research methodology of CAD system on NIR and MRI images, the tissue samples will first be normally evaluated by experienced pathologists to be separated into the following class labels.
(1) Stroma: STR (normal muscular tissue); (2) Benign prostatic hyperplasia: BPH (a benign condition); (3) Prostatic intraepithelial neoplasia: PIN (a precursor state for cancer); (4) Prostatic carcinoma: PCa (abnormal tissue development corresponding to cancer). Figure 1 demonstrates the NIR and MIR multispectral acquisition system. In this system, a liquid crystal tunable filter (LCTF) (VariSpec) is plugged in the optical path in the middle of the chilled charge-coupled device (CCD) camera and the light source. e LCTF has a bandwidth accuracy of 0.25 nm to 20 nm. e wavelength is controllable over several spectrums.
is allows capturing a different MIR multispectral images of the tissue samples at diverse spectral frequencies. Usually, the experiments are carried out on varying numbers of bands to illustrate the impact of multispectral imaging.

Challenges of Using NIR and MIR Images.
One of the challenges of using these images is interpreting the imaging data to reflect cancer margin accurately. It is difficult to obtain accurate, reproducible, intent, and precise assessments in cancer research. e dilemmas occurred because of the biological dissimilarity, variability of personnel, and natural unpredictability. e NIR imaging technique characterizes cancer by adding fluorescent information to each pixel in the acquired image. For prostate cancer, the NIR imaging CAD system could be analyzed and classified to construct several sharp diagnostic rules.

Magnetic Resonance Imaging (MRI).
e PCa representation in the past decade has changed substantially. Nevertheless, the mp-MR of prostate was mostly confirmed by the increases in magnet strength between 1 and 3 Tesla. e mp-MR is used for anatomic and functional imaging techniques: the anatomical imaging includes T1 or T2weighted stimuli, and the dynamic contrast-enhanced (DCE) imagery is included in functional imaging. An MRI of 1.5 Tesla is currently recommended with an endorectal coil or an MRI of 3 Tesla without using an endorectal coil. e MRI scan of multiples parameters is one type that produces accurate photographs of the prostate rather than a normal MRI scan, because it combines up to four different image types. From the mp-MR, the doctor can conclude whether or not cancer is present in the tested prostate and how rapidly cancer is expected to develop. e combination of different images of the MRI provides vast quantities of data. Supporting techniques or methods like CAD is vital for a fast, accurate, and reliable clinical decision. e PCa diagnoses are necessary to interpret MRI images for prostate radiologists, and this knowledge is not widely available. Intelligent CAD implementation will significantly increase the efficiency of less experienced PCa diagnosis observers.
mp-MR Acquisition. In 2012, a discussion was raised up for the use of a combination of images (T2W, DWI, MRS, and DCE) and practical methods in the Prostate Imaging Reporting and Data System (PI-RADS) on prostate MR guidelines Barentsz et al. [30]. In comparison, each type of these images has specific points of strength that will help us in the diagnosis process. For the anatomical structure of prostate gland, T2W images are the ideal choice; DWI and MRS add the characterization of lesions, while DCE-MR imaging has a high sensitivity to cancer detection. In PIRADs methodologies, T2W, DWI, and DCE are the elementary components used during the mp-MR test [31]. e DWI is the central sequence of determination for the PZ. For the TZ, and the T2W is the primary sequence of judgment. A combination of anatomical and functional imaging is required in clinical practice to achieve high and reliable results. Recent studies have shown an increased interest in CAD systems development to detect PCa on the mp-MR approach and its representation [32][33][34][35]. In mp-MR CAD devices, T2W-MR images are commonly used. e power of multiparameter MR comes from gathering the strength points of T2W, DWI, and DCE-MR types. erefore, this type is the most widely used currently.
T2WI and T2 Mapping. Taking the anatomical structure of the prostate with a T2W image is an effective way to obtain the ideal acquired image. e first and the most important step in an mp-MR protocol is the acquisition phase for a high-resolution T2W prostatic image. In T2W images, hyperintense signals are present in the peripheral zone PZ of the prostate, while a low signal exists in the central area and on TZ, which allows an easy and straightforward recognition for the defined zonal prostate anatomy.
e Peripheral Zones of the PCa usually are represented as a low-signal region in T2W. Nevertheless, the pattern of tumor existence and growth will affect the appearance of the prostate anatomical structure. In [36,37], images in T2W-MRI type for prostate were used as a specific technique for PCa detection in the transition zone. Due to the reliability of this type, it has been used for determining the area of interest of the tumor. At the same time, T2W-MR images contain more dark pixels than luminous pixels in the tumor area of interest, whereas the normal tissue ROI has more luminous pixels than dark pixels. A CAD system can use different characteristics, including fractal elements, textural features, and signal strength. As PCa and PZ usually have different textures in T2W-MR images [38] and since mp-MR can be challenged in the detection of TZ cancer [39], a CAD system can help identify irregular lesions that can examine feature-based positions of the lesion. e device can also be used to diagnose Transition Zone cancer. e quantitative T2 values are given by T2 maps images. Because of the requirements of traditional T2 mapping methods, such as excessive scanning times for multiple single eco-acquisitions with a variety of echo-time settings, the T2 mapping in most clinical applications is not supported. Several new sequences have recently been used to speed up T2 quantification [39][40][41][42]. e T2 values of the malignant tumor areas that had been histologically demonstrated were significantly lower than suspicious nonmalignant lesions or the standard areas [40]. Quantitative T2 assessment increases PCa detection accuracy and/or sensitivity [43], and the evaluation of aggression can be incorporated into CAD systems [44,45].
DCE-MR Images. DCE image is an essential oncological tool for defining prostate tumors. Indicators referring to signal strength should be used; because the semiquantitative DCE-MRI results are proportional only to the patient, and the baseline amplitude depends on the patient and the MRI imaging guidelines. DCE is so sensitive to the changes of vascular, blood flows, or extra-cellular space. So, it is practical to use quantitative or semiquantitative techniques to analyze and measures the changes of vascular, blood  e clinical application of DCE-MRI in prostate cancer is differentiating between malignant lesions and healthy tissues by enhancement and contrast agent washing out.
e Pharmacokinetic (PK) compartment model is one of the most commonly used quantitative analytical approaches of DCE-MRI imaging. PK allows the generation of the PK parameters, such as K-trans transferring gadolinium contrast from the vasculature to the tumor, which reflects forward vascular infusion and permeability of the contrast uptake and washout [46] and the methodology for estimating input functions, which can have significant effects on the PK model-ling parameters measured [47]. e experimental findings from a DCE-MRI acquisition can be replicated by an analytical approach based on phenomenological universities [48,49]. To analyze DCE-MRI data, several CAD systems were developed. For example, Vos et al. [50] have developed a CAD model that can differentiate PCa tissues and nonmalignant Peripheral Zone tissues. Moreover, the system has achieved 83% (0.75-0.92) as an average accuracy. An automated segmentation model has been developed to improve the accuracy of CAD diagnosis [51] to produce five-level scores for checked images in DCE-MRI and T1 images. ere were many attempts to construct and design an automated prostate cancer diagnosis system. Puech et al. developed a prostate CAD system [52,53]. Typically, DCE-MRI requires enough acquisition time to produce a good resolution image.
In contrast, DCE-MR has lower spatial resolution than other sequences, mainly, if DCE-MR is carried out quickly over a short period. Limitations to DCE-MR interpretation involve duplication of enhancement features between benign and malignant TZ areas. ere is also major hypervascularity in benign prostatic hyperplasia and other benign inflammatory disorders within the TZ [54]. Contrast-enhancement diagnostic models have reduced performance when used in zones, so zone-specific models should improve mp-MR PCa classification [55].
Diffusion-Weighted MR Images. In general, the DWI is generated with different b-values to measure the apparent diffusion coefficient (ADC), and the ADC is shown as ADC map for each pixel of the image. e spread of water molecules in tumor tissue reflects the architecture of tissues, such as cell density and the nucleus-to-cytoplasm ratio and ADC values. As a predictor of Gleason score in PCa [56,57] for these reasons, the ADC values have been given attention. Data from DWI trials have shown that tumor aggressiveness can be detected using type DWI [57][58][59].
Bi-Parametric MR Images. e drawbacks of DCE imaging, which are the expensive contrast agent management, long scanning period, have led to the use of new types of images. A biparametric MR (bp-MR) protocol, which requires only T2W and DWI, is being widely tested for PCa diagnosis due to the rapid rise in the use of mp-MR for PCa. e biparametric MR (bp-MR) protocol is systematically used for PCa treatment, including T2W and DWI. e protocol of bp-MR can be executed of costly contrast medium for 15 minutes while retaining sufficient diagnostic reliability, all of which may facilitate the use and increase the efficiency [60,61]. Currently, many types of research focusing on bp-MR or comparisons between bp-MR and mp-MR were published [61,62]. In an article published in 2017, Greer et al. [63] suggested that the application of DCEI-MR could benefit the diagnostic of prostate cancer, because abnormal DCEI-MR findings improve the cancer detection rate in each of the PI-RADSv2 categories 2, 3, 4, and 5. However, those who advocated using bp-MR or opposed the use of DCEI-MR suggested that compared with mp-MR.
Magnetic Resonance Fingerprinting (MRF). Magnetic Resonance Fingerprinting (MRF) is the newly invented mapmethod for nuclear magnetic resonance imaging introduced by Ma et al. [64] that could concurrently multiple approximate parameters of quantitative tissue properties, such as T1 and T2 relaxation times by transitory signal evolutions and data analysis. With this technology, MRF can provide a solution to this problem in effective and fast scan times by collecting quantitative measurements. Previous studies have shown that MR fingerprinting parameter maps have strong reproductivity in firm tissues in the supratentorial human brain region [65]. e support of larger multicenter, randomized control trials is still necessary if reproducibility and repeatability can be enhanced in PCa detection. Furthermore, 3D-1H-MRSI was also shown to be reproducible by Lage-maat et al. [66]. However, it was suggested that the reliability had been constrained by a small sample; further population-based experiments were also needed for proof of this in the future.
Challenges with Using MR Images. Despite the enhancement of the accuracy for PCa detection with mp-MR in the diagnostic process, the widespread use of mp-MR for PCa diagnosis has been embedded by various challenges.  Figure 3: Flow chart of a short-term Tabu search [82].

Mathematical Problems in Engineering
Standardization. e features of the used scanner determine the output quality of Prostate Cancer detection using mpMRI (vendor, strength of the magnet field, technique, technology, etc.), patient factors (his movement, preparation etc.), and most significantly, the radiologists' interpretation [67].
Personal Challenges, Inter-and Intra-Reader Variability. Even with standardization of diagnosis and interpretation, there are certain drawbacks of perception, such as harmless circumstances, ERC-related artifacts, or changes in appearance the following care. High-level prostate mp-MR expertise is crucial to correct management that is not accessible in many centers [2,68]. e variance of inter-reader is also a problem for expert readers [69,70].
e Cost. Multi-Parametric MR (mp-MR) is a rather costly analysis, or the costs are high.
e mp-MR's ability to prevent biopsies, reduce overtreatment, and contributing to a higher quality of life can result in overall cost-effectiveness, but further studies are needed to confirm this [25,71]. Reducing the mp-MR protocol to bp-MR could reduce the scanning time from 40 minutes to 15 minutes, avoid using the contrast medium, and reduce costs [60].

Proposed Methodologies of CAD System
We propose two research methodologies of CAD systems based on NIR, MIR, and MRI images in this work. We describe both methodologies in the following subsections.

Research Methodology of CAD System on NIR and MIR Images.
is section describes the research methodology used on NIR and MIR images for designing an intelligent CAD system in detail. Figure 2 illustrates the feature extraction, infrared-bands feature selection, and classification phases for prostate cancer diagnosis. Mainly, we adopt a traditional pattern recognition approach, which involves the following stages: (1) Inputs: e methodology supposes that the system receives the VIR, NIR, and MIR multispectral images from stroma samples, BPH, PIN, and PCa as inputs. ese images are also used to measure and evaluate NIR and MIR technology's performance as benchmarks.
(2) Feature extraction: is stage is extremely crucial in extracting and obtaining relevant features for each band. e system development will be deployed to create three main phases: (i) feature extraction phase, (ii) feature selection from useful bands, and (iii) classification phase. After obtaining multispectral images, important features are extracted from multispectral bands images. en, the extracted features from useful bands are selected. Finally, the selected features are classified.

Feature Extraction.
After multispectral image acquisition, essential features are extracted from NIR and MIR images. For this, three effective algorithms are used, which are the traditional first-and second-order Haralick texture features [72], the histogram-based features such as local binary pattern (LBP), and local phase quantization (LPQ). e following subsections describe these algorithms in detail.

NIR and MIR Local Binary Patterns (LBPs)
. e LBP was first introduced by Ojala et al. [73] as a gray-scale invariant pattern to measure texture feature in the source images. Its tolerance for any illumination changes characterizes it. Moreover, it is effective compared to other approaches, so that the images can be analyzed in real-time. Generally, the LBP operator is derived from the related information of texture on the local neighborhood. LBP can be extended to multispectral images and easy to handle multiscale analysis, which is quite efficient for prostate cancer classification applications.
NIR and MIR Local Phase Quantization. Local phase quantization (LPQ) [74] is another attractive and powerful texture descriptor for many real-world applications, such as image retrieval and face recognition. Its higher accuracy and computational complexity characterize it. LPQ uses the local phase information, which is computed in a window for every image position. LPQ is similar to LBP in its support for multiscale analysis and the ability to extend to multispectral image analysis.

Infrared-Multispectral Band Selection.
In this work, we proposed the use of a considerable amount of NIR and MIR bands. For this regard, band selection is crucial as it is challenging to figure out which bands have the relevant information, or which two bands are redundant. Selecting the essential bands helps reduce the cost of feature extraction in test images. We dealt with band selection as the feature selection problem from the mathematical point of view. However, instead of handling the features, we search for the optimal number of bands, where some of the current algorithms for feature selection can be exploited.
Let n be the number of multispectral bands extracted acquired from the NIR and MIR spectrum. We used sequential forward selection (SFS) [75] as landmine detection using multispectral images [76]. SFS is known to be a simple greedy sequential search algorithm. Another algorithm is the sequential forward floating search (SFFS) and sequential backward floating search (SBFS). Compared with SFS, the SFFS and SBFS algorithms are more efficient, especially finding optimal solutions for minor and medium-scale problems [77]. Nevertheless, SFFS and SBFS algorithms cannot converge to locally optimal solutions for large-scale issues when n is more significant than 100 [78,79]. Recently, TS and genetic algorithms are used profusely to handle the problems of having an exponential and noisy search space with numerous local optima [79][80][81][82]. In this work, we aim to introduce the Tabu search for band selection. Tabu search is described in the following subsection.
Tabu Search: Tabu search was introduced by Glover [83]. Tabu Search (TS) can be described as a meta-heuristic, which is based on three principles: first, the use of data structure to memorize the history of the evolution of the search process; second, the use of a control mechanism to strike a balance between the acceptance, or not, of a new configuration, based on the information recorded in the taboo queue regarding the desired restrictions and aspirations; third, the intensification and diversification strategy procedures using short and long term memories. Accordingly, the TS algorithm is recommended to be used in the research methodology. e TS algorithm is given in Figure 3. [82,[84][85][86][87] have been proposed for performing the classification task. In this section, we focus on investigating the methods for improving classification accuracy. e literature has already resulted in effective classifiers using a round-robin method, in which multiclass complex problems have been broken into many binary-class simple problems. Investigating the new classification techniques, such as logistic regression, Dempster-Shafer theory, decision template, and fuzzy integral method, is highlighted to determine each band's contribution of the NIR/MIR to the overall classification process. ese new classification techniques are introduced below.

Logistic Regression.
e logistic regression method transforms the classification problem into a regression problem [88,89]. e process tries different performances of distinct classifiers. It seems like a Borda count method in terms of generalization and assigning optimal linear weights to the rank scores for reflecting the significance of the separate classifiers in the combined decision.
Fuzzy Integral. For the diversity and capacity of different classifiers aggregation, the fuzzy integral has been applied to classifiers combinations [90,91]. A fuzzy measure is utilized to reflect the importance of classifiers from a complete set of learners regarding the class. Choquet fuzzy integral and Sugeno fuzzy integral are two essential types of fuzzy integrals. A fuzzy measure of a classifier is computed from several fuzzy densities to represent the particular importance of the classifier. e fuzzy measure can be consistent with these fuzzy densities. e classification result of a class c is a cooperation between the evidence represented by the class c in the different classifiers and competence defined by the fuzzy measure. e fuzzy measure vector might be specific for the current data sample and new for each class. If the ordering of the classifier support is the same, the two vectors of fuzzy measure will be only the same.
Decision Templates. In this type of supervised learning method, the classification results of distinctive classifiers for a given data test set can be arranged in a matrix called a decision profile (DP). e elements of the DP matrix (c, l) are a classifier c for class label l. e decision templates (DTs) combine the classifiers [92,93], in which each decision template is computed from a test example given a trained classifier. A decision template of a class is the centroid of this class in the intermediate of the feature space. e anticipated decision profile of this class is related to this decision template. e prediction of a class label in a combination of classifiers can be calculated using a similarity measure (e.g., Euclidean distance) between the existing DP (of the test example to be classified) and the DTs classes.

Dempster-Shafer Combination.
e Dempster-Shafer calculus [94,95] is a method used for controlling the degrees of belief. Compared with the Bayesian approach, it is more general and does not involve adding additive probability. In [93], the author introduced a method that uses the theory of the Dempster-Shafer's evidence to combine the outputs of multiple classifiers and give a soft production. e training process of this approach is equal to the training of the decision templates approach. However, rather than computing the similarity between the decision profile and the decision template, the new data points are classified by calculating the proximities between each classifier output and decision template. For every classifier and every class, two basic probability assignments (BPAs) are constructed based on the proximities. By using the BPAs, the degrees of belief are computed based on the orthogonal sum rule. From the degrees of belief, the final degree of predictions is calculated for each class.

Research Methodology of CAD System on MRI Images.
e use of MRI technology to diagnose PCa requires radiologists to interpret vast amounts of images and involves knowledge of experience that is not widely available. Automatic methods may simplify the radiologists' job, increasing read time and variance of the reader [62]. Automated approaches have been established to help less experienced readers reach the same performance level as professionals. Figure 4 shows a typical workflow of a computerized prostate cancer diagnosis system using mp-MR images.

MR Acquisition.
e mp-MR analysis typically consists of an anatomical sequence (T2-weighted (T2W)) images and multiple functional sequences, generally DWI and DCE sequences. e sequence choice is based on the medical need and time and costs limitations [61]. ere is currently growing research into the use of mp-MR as a triage tool. It is unlikely that patients with negative mp-MR will have any clinically meaningful PCa and could avoid biopsy. When using mp-MR as a triage test, 27% of patients are diagnosed Mathematical Problems in Engineering with a biopsy, and less than (5%) are diagnosed with clinically insignificant PCa [94][95][96]. erefore, young patients with elevated or reduced PSA must retain the normal TRUS + B until a definitive conclusion is reached on the negative predictive value of mp-MR [97].

Preprocessing.
Preprocessing the images requires standardization of image intensities, where the T2W image sequence, in particular, suffers from inter-and intra-patient variability, even for images collected using the same scanner and procedure. Other standard methods of preprocessing include noise filtering and correction of field bias [24]. e selection of steps to preprocess depends on the dataset and the software.

Registration.
Registration is known as a method of aligning two or more images, may be helpful during the evaluation to compensate for patient movement and changes in bladder rectum filling. Long-term MR test procedures (e.g., DCE imaging) increase the likelihood of substantial patient movement and identify images [3,15].

Segmentation.
e segmentation process is a critical stage of the diagnosis process. e accurate diagnosis depends on the precise prostate segmentation from the obtained image [98][99][100]. e lack of clear boundaries and considerable variability in the prostate shape and appearance makes manual delineation a challenge. It is well known that the T2W image sequence offers the best analysis of prostate anatomy and the ability to delineate margins and distinguish between the zones of the prostate gland [101]. e manual delineation is very time-consuming and requires prostate MRI experience. In the literature, automated methods include map-based methods, model-based methods (e.g., effective form model), and edge-based methods [102]. In recent years, techniques have focused on the use of deep learning approaches, especially deep convolution neural networks (CNNs), that have made significant progress in the analysis of clinical images, including prostate segmentation [103]. Recently, more emphasis has been found on the segmentation techniques for prostate zones. Many PCa is located at the Peripheral Zone PZ, and since the PCa's biological activity varies between zones, this information is essential for making medical decisions [62,[104][105][106]. Recent zonal segmentation studies have used various voxel identification approaches (3D pixel analog) and active form models [107][108][109][110][111][112][113][114][115]. e Current first place in the MICCAI Grand Prostate MRI segmentation challenge (PROS-TATE12) is a CNN solution (reaching a coefficient of 0.8721 Dice score) [109]. One of the main challenges in zonal segmentation is the absence of gland apex and base characteristics and the gradients [106,116].

Detection.
e beginning was in 2003 with Chan et al. [117], the initial work on automated methods in prostate mp-MR focused on identifying suspect areas for directed guided biopsies of MR. e most common approach in the literature is to identify voxels as either PCa or normal tissue based on various image characteristics such as texture, signal strength, and data on gradients. A study conducted by Rampun et al. examined 215 T2WMR texture features for classifying PZ voxels as malignant and benign using 11 different classifiers (e.g., support vector machine (SVM), random woods, naïve Bayes, and k-nearest neighbor) [118]. Combining the T2W sequence with one or more usable arrangements provides better detection over a single image mode. Extracted object characteristics from T2W, DCE and DWI resulted in AUC of 0.95 in a study by Peng et al. with using a linear discriminatory analysis to classify interest regions as either cancer or normal [67]. Nevertheless, most studies use T2WMR in conjunction with DWI, including ADC and/or DCE scanning. MRS was also investigated. Probably because of the difficulty and duration of data acquisition, the MRS has not achieved broad acceptance [119]. Various studies agree that a zone-aware Classifier significantly improves PCa [78,120] identification. Many reported PCa detection algorithms to record an area between 0.80 and 0.89 under the receiver characteristic curve (AUC) [15]. One of the research showing the highest performance in literature is the analysis described by Peng et al. [67].

Classification.
e selection of treatment for patients with PCa is based on clinical factors such as PSA, GS, age, and comorbidity. As previously mentioned, the GS is the most important indicator of disease progression, mortality, and outcomes. Since the GS of prostate biopsies sometimes varies from the true GS of RP, there is a medical need to distinguish slow-growing, indolent PCa better from those of RP with catastrophic outcomes of clinical significance [119]. e mp-MRI is usually used to determine PCa aggressiveness for noninvasive, pretreatment purposes. GS and ADC values are substantially associated with lower ADC values suggesting higher GS. Moreover, other studies found an association between DCE Parameters, frequency of T2W signal, and aggressiveness of PCa. Nevertheless, these single parameters are not enough to predict the GS alone [121][122][123][124][125][126]. Many studies have investigated algorithms with multiple image characteristics, such as shape, T2W, DWI and ADC frequency, to differentiate malignant from benign lesions or to classify lesions into clinically insignificant (GS6) or clinically insignificant. (GSB7) promising results [127][128][129][130][131][132][133]. Holtz et al. studied a small, medium, and high- grade three-class classifier compared to a two-class system and showed low performance for the three-class system. One analysis achieved accuracies of up to 0.93 for two-class classification of GS range versus GS range 7, and 7 (3 + 4) versus 7 (4 + 3) using ADC and T2W derived feature [132]. e sensitivity of 100 percent and specificity of 76.92 percent has been achieved in a more recent study focused on a multimodal convolution neural network to distinguish GS6 from GS MUSK7 [129]. Because the prognosis and therapeutic options vary for each GS category, it would be of medical interest to determine lesions more accurately into more than two or three classes of clinical interest.

Diagnosis.
ere is a high clinical need for developing fast and cost-effective computer diagnostic tools. Researchers have focused on this area, and there were plenty of attempts to build CAD automated systems for diagnosing prostate cancer using NIR, MIR, and MRS classification and supportive imaging types such as DCE-MR analysis. To increase diagnostic precision, the MR anatomic images (T2W) are combined with functional techniques because the single method cannot adequately detect and characterize PCa [119].
Chan et al. have one of the first attempts for implementing an mp-MR CAD system for the diagnosis of PCa using the approach of T2, T2W, and line scan to detect the predefined area of PZ of the PCa [117]. In the field of fuzzy logic techniques, Liu et al. [134] presented PCa detection using multispectral MR and NIR images. Moreover, Tiwari et al. [135] used the MRS with T2W-MRI to identify the presence of PCa by the existence of voxels affected by PCa.
But in the research work [67], Peng et al. combined 10th percentile ADC, average ADC, and T2W skewness to differentiate between normal and abnormal tissues of the prostate. In their approach, they obtained an AUC value of 0.95.
Litjens et al. [136] developed a fully automated CAD system to recognize prostate cancer by performing two phases. e first is the detection phase that includes segmentation of prostate on T2W MRI, extracting features of the voxel, finally classification and select a candidate. After that, the second phase is responsible for performing processes on the output of phase one. In phase two, segmentation and feature extraction and classification for the candidate. Although the use of 347 patients data, the system reaches an AUC value of 0.889.

Aggressiveness.
It is important to construct an intelligent system for differentiating between high-grade cancer and indolent, slow-growing tumors based on initial PSA level, Gleason score, and patient age [137,138]. One of the commonly employed statistical variables of PCa is the Gleason test, which utilizes a ranking system to determine the classification in PCa. Tumors of higher grades tend to grow rapidly and spread faster than tumors of lower grades.
Research by Yamamura et al. [139] found that the ADC values and the Gleason level were very strongly negatively correlated, although MRS displayed no significant correlation. PCa aggressively test approaches include noninvasive DWI, DCE-MRI, and MRS. Oto et al. [140] stated that the multi-parametric MR such as MR spectroscopy, T2weighted imaging, dynamic contrast-enhanced MR (DCE-MRI), and diffusion-weighted imaging (DWI) is the best modality of imaging for staging and diagnosis of PCa. In vivo, MRS imaging showed an improvement in the Gleason level [141,142] ratio (choline + creatine)/citrate. e magic angle spinning ex vivo (HRMAS) MRS [143] has also shown this relation. Nevertheless, there was no link between metabolite ratios and aggressiveness [144,145] observed in other in vivo MRS imaging studies.
Changes in the signal intensity were associated with its aggressiveness [146] on the T2W-MRI for PCa detection. Lee et al. [147] demonstrated the possibility of improving the prediction of minor PCa in candidates for active monitoring treatment through simple measurements for the diameter of suspicious tumor lesions on DWI.

Biopsy Guidance.
e biopsy may not accurately depict the whole gland, but examines a small portion of the prostate. e Gleason value in a random systemic biopsy, driven by TRUS, is usually believed to lower the test, since the biopsy may not have shown a less distinct trend [148,149]. Natural routine biopsies of TRUS also require multiple biopsy procedures combined with pain and possible disease [150]. Lesions have to be reliably identified, described, and treated during biopsy for minimizing overtreatment and amount of biopsies. In order to improve the detection rate of prostate biopsies, more successful imagingled guided biopsy approaches are being studied. In recent years, many minimally invasive, oriented approaches to conserve organ were used as additional options in contrast to the revolutionary use of PCa [150]. e following table (Table 1.) illustrates the studies in this field.

Discussion, Observation, and Future Directions
CAD diagnosis approaches to prostate cancer from medical imagery are based on various widely used NIR, MIR, MRI and TRUS dissecting methods. ese are highly invasive and can assist in picture-orientated operations. ey aim to detect prostate cancer early and can be used to enhance the remediation of patients. Nonetheless, problems remain that must be overcome to be used in the clinical environment so that a more efficient diagnostic method can be used. ese include the following: ( proximity of adjacent nonprostatic tissue (e.g., the bladder), (iii) in many cases, DWI plays a major role in medical imaging. Applications such as this are to build an effective CAD imaging system for the detection of prostate cancer. is paper offered a summary of these systems covering noninvasive CAD systems from DWI. In this paper, we discussed current approaches for each step of machine-based learning CAD systems, with a focus on their advantages.
(3) Construct a successful and efficient, and intelligent CAD system that will reduce prostate cancer deaths by an earlier disease diagnosis. Researchers should then establish and test approaches to prostate diagnosis to enhance their results.

Conclusion
Since the 90s, they are constructing CAD systems have been recognized as one of the hottest fields for researches. Medical Precise and effective prostate cancer detection systems play a significant role in managing the mortality rate of serious malignancy among males. is paper presented an in-depth review of current computer-aided methods of a prostate cancer diagnosis. Mainly, the NIR, MIR, and MRI techniques have been discussed, which are used in the characterization of prostate tissue. In theory, the features of tissues by the NIR, MIR, and MRI images are essential for giving promising results of PCa detection, despite modeling interactions between these images and tissues. It is a significant achievement, though, with the inclusion of automatic ultrasound tissue characteristically methods, which are needed to correct pathological conditions with the acoustic properties of the tissue. It should be noted that intelligent computer-aided diagnostic systems are intended primarily to offer a second opinion and are not intended to replace a doctor. e rule also applies to prostate cancer. e occurrence of false-negative results, in this case specifically, means that a different perspective can be advantageous. If the biopsy is possible with this helping opinion, the biologist will select the biopsy sites.

FMT:
Fluorescent molecular tomographic NIR: Near-Infrared MIR: Mid-Infrared MRI: Magnetic resonance imaging PSMA: Prostate specific membrane antigen FMT: Fluorescent molecular tomographic AUC: Area under a receiver operating characteristic curve FP: False positive FN: False negative Sp: Specificity USPSTF: United states preventive services task force FLD: Fisher linear discriminant CI: Confidence interval csCPa: Clinically significant prostate cancer FLD: Fisher linear discriminant OR: Odds ratio PSA: Prostate-specific antigen CRF: Conditional random fields ROI: Region of interest ADC: Apparent diffusion coefficient CAD: Computer-aided diagnosis DCE: Dynamic contrast-enhanced LDA: Linear discriminant analysis LLE: Locally linear embedding MRS: Magnetic resonance spectroscopy NLDR: Nonlinear dimensionality reduction PCA: Principal component analysis Benign prostatic hyperplasia PIN: Prostatic intraepithelial neoplasia.

Data Availability
No data were used to support the study.

Conflicts of Interest
e authors declare no conflicts of interest.