White-light cystoscopy (WLC) is commonly used for BCa detection and follow-up after initial treatment of NMIBC. Despite its availability, WLC is affected by low sensitivity, mainly due to its operator-dependent nature. One of the most difficult challenges is to distinguish flat lesions from misleading mucosal non-specific reactivity [25]. Shkolyar et al. proposed an image analysis platform based on convolutional neural networks (CystoNet); they showed a per-frame sensitivity and specificity of 90.9% and 98.6%, respectively, in 54 prospective cases, resulting in three on three correct diagnoses of flat tumors on the final pathology report [26]. Wu et al. described a cystoscopy artificial intelligence diagnostic system (CAIDS) as being quicker and more accurate than expert urologists in the diagnostic assessment. With a latency diagnostic time of 12 s and a high accuracy evaluated as an area under the curve (AUC) of 0.94, the authors claimed to improve the detection of commonly misdiagnosed cases such as flat carcinoma in situ (cis) [27]. Recently, Yoo et al. presented an AI-enhanced platform able to predict BCa grading based on tumor color with the red/green/blue (RGB) method; the performance was ≥98% for the diagnosis of benign vs. low-and high-grade tumors and >90% for the diagnosis of chronic non-specific inflammation vs. cis compared to conventional WLC [28]. Mutaguchi et al. aimed to reduce the risk of early recurrence due to overlooking of tumors during endoscopic resections by proposing a diagnostic system (Dilated U-net) trained on 1790 cystoscopy images categorized by the pathological T score from 120 patients who underwent TURBt [29].

In the context of imaging-enhanced visualization, blue light (BL) photodynamic diagnosis (PDD) provides better diagnostic accuracy and more complete tumor resection at the time of TURBt, potentially reducing BCa recurrence [30]. However, controversy exists about its impact, which was recently highlighted by Heer et al. within an open label randomized clinical trial including 538 NMIBC patients [31]. Thus, novel emerging AI-based algorithms have been advocated to refine such a diagnostic pathway. Ali et al. proposed an AI platform able to predict malignancy, invasiveness, and grading from BL cystoscopy images. The results showed a sensitivity and specificity of 95.77% and 87.84%, respectively, for BCa diagnosis, while the mean sensitivity and mean specificity for tumor invasiveness were 88% and 96.56%, respectively [32].

Notably, AI technology displays interesting potential as a learning tool, which may improve urologists’ performance and cystoscopy skills depending on their experience level [33]. Ikeda et al. concluded that physicians’ diagnostic accuracy can be objectively evaluated using their GoogLeNet platform since its detection accuracy was comparable to the one of an expert urologist [34]. The most relevant limitation of AI-enhanced cystoscopy imaging is the limited availability of the ML platforms. To address this issue, Du et al. presented the EasyDL platform designed as an application for mobile phones. This system showed an accuracy rate of 96.9% in BCa detection, which encompasses, at the same time, significantly more manageable possibilities of utilization as the software provides an output based on cystoscopy images uploaded through common smartphones [35].

In summary, AI applied to cystoscopy assessment attempts to fill the gap due to interobserver variability and to reduce the risk of misdiagnosis, especially when non-univocal findings are detected during routine cystoscopy. Data from the recently reported studies showed promising results in terms of accuracy; however, the use of AI is still limited by the low availability of AI platforms, which does not allow the application of these systems in current clinical daily practice.

Current International guidelines highlight the role of computed tomography (CT)-urography scanning and multi-parametric magnetic resonance imaging (MRI) in BCa cT staging [36]. AI and DL platforms are well applied to radiological images re-elaboration tasks, ensuring enhanced accuracy in radiological diagnosis across different settings and clinical scenarios [37]. In a retrospective study on 441 BCa patients, Zhang et al. aimed to validate a DL model able to preoperatively predict BCa invasiveness status by analyzing CT images. The authors divided the study population into development, internal validation, and external validation cohorts. The performance of the model was compared to the individual subjective assessment of two different radiologists. The model showed a relatively good performance in all cohorts and outperformed the two radiologists regarding the accuracy, reaching a sensitivity of 0.733 and a specificity of 0.810 in the internal validation cohort and a sensitivity of 0.710 with a specificity of 0.773 in the external validation cohort. Of note, the model demonstrated a lower sensitivity compared to the radiologists. As a limitation, the authors noticed that the DL model considered tumor size above 4 cm as the key feature to detect muscular invasion, potentially leading to misdiagnosis in some cases [38]. Similarly, Yang et al. developed a DL convolutional neural network with the aim to distinguish MIBC and NMIBC on CT scan frames. In this study, eight algorithms were tested and exhibited an AUC ranging between 0.762 and 0.997 [39]. Liu et al. elaborated a DL model intended for the prediction of BCa stage before surgery; according to the authors, the sensitivity rate of DL diagnosis was 94.74% notwithstanding CT detection drawbacks such as poor spatial resolution; therefore, there was a need to combine more frames to obtain an adequate positioning [40]. Taguchi et al. prospectively applied DL technology to improve the T2-weighted MRI frame output of new generation 3T MRI. Denoising DL-mediated processing of raw MRI images has been proven to enhance the accuracy of the VI-RADS score calculation for BCa detrusor invasion [41]. Yu et al. have recently proposed their Cascade Path Augmentation Unet (CPA-Unet) based on T2-weighted MRI frames, which could elaborate proper segmentation of bladder wall layers, identifying the depth of local tumor infiltration [42]. An interesting application of ML applied to CT scans was described by Cha et al. [43]; they proposed a CT-based computerized decision-support system for MIBC patients undergoing neoadjuvant chemotherapy (NAC). The ML-platform showed enhanced performance in identifying patients who would have achieved a complete response to NAC by analyzing post-therapy CT scans.

To summarize, AI, particularly DL, could play an interesting role in enhancing the radiological diagnosis of BCa. However, DL models still suffer from a lack of satisfactory sensitivity as well as from some intrinsic limitations of the systems, as such, the identification of the radiological features necessary to produce a correct diagnosis. Moreover, many studies are affected by the limited number of patients enrolled. In light of the above, despite several encouraging results, further studies are pending in order to produce more robust evidence allowing the stable introduction of these DL systems.

Due to its prognostic importance, pathological evaluation of histological samples plays a key role in BCa management. Of note, interobserver variability may lead to incorrect pathological interpretation [16]. AI technologies may provide an integrative tool in this scenario, aiming to enhance interpretation reproducibility through a semi or fully automated slides reading procedure. Jansen et al. [44] proposed an automated detection and grading network for NMIBC based on DL technology. The authors showed a correct grading of 76% low-grade and 71% of high-grade BCa, according to the consensus reading. Chen et al. proposed a ML-model able to develop automatic diagnostic and clinical prognostic models based on histological samples, displaying high accuracy rates. Despite these promising results, the authors acknowledged limitations of the retrospective design and the lower accuracy of ML diagnosis compared to traditional diagnosis performed by an experienced uro-pathologist [45]. Yin et al. developed a model able to distinguish between Ta or T1 features on sample images using six supervised learning methods (91–96% accuracy) [46]. AI-enhanced histological evaluation could play a role also in MIBC setting. Harmon et al., in a study including 307 patients, proposed an AI-enhanced model based on features of hematoxylin and eosin-stained RC specimens slides, which was not able to predict the risk of lymph node metastases [47].

Substantial work by multiple groups has defined key molecular subtypes of UC characterized by distinct gene signatures, varying expression of potential drug targets, and differing therapy sensitivity [48,49,50]. In this context, AI-enhanced molecular analysis is another innovative frontier in detecting mutations in key molecular pathways potentially providing tailored management in BCa systemic therapies, including conventional chemotherapy and novel targeted therapies. For instance, the literature has produced robust evidence on the role of fibroblastic growth factors receptor (FGF-R) pathways in BCa tumorigenesis and progression [8,13,51]. Nevertheless, mutational status detection still demands genetic sequencing techniques. Loeffler et al. and Velmahos et al. described an ML algorithm able to identify FGFR2/3 mutational status based on AI-assisted analysis of diagnostic hematoxylin and eosin-stained BCa slides, which might improve the suitability of FGFR inhibitor administration [52,53]. In addition, Xu et al. developed a model able to identify a specific AI-derived gene signature (AIGS) for predicting the therapeutic response or providing prognostic information for individualized follow-up. However, some limitations have been acknowledged such as the retrospective design [54].

Considering these findings, AI also displays interesting perspectives in pathological and molecular evaluation. A recent review on this topic has been published by Wessels et al. [55]; the authors included 16 studies regarding AI-enhanced hematoxylin and eosin-stained slide analysis, focusing not only on BCa, but also on upper-tract urothelial carcinoma, prostate cancer, and renal cell carcinoma. One interesting perspective described by Wessels et al. is the potential ability of AI-trained models to detect new and still unknown histological patterns with prognostic significance. This new scenario can potentially lead to an improvement in risk stratification and oncological outcomes; however, the prognostic significance of these patterns is yet to be established.

On the other hand, larger datasets are required to train DL models to enable them to produce a satisfactory performance; this could turn into a thorny issue when considering rare diseases such as variant histology BCa [46,56]. Many of the previous DL models were set up using only specimens from the primary tumor, while including samples from metastatic sites as well could eventually enhance the accuracy of DL-assisted diagnosis [53]. Moreover, many studies still suffer from a retrospective design and a limited cohort. Thus, AI needs further validation in the field of histological and molecular pathology before entering into routine clinical practice.