Bladder Cancer and Artificial Intelligence Emerging Applications

INTRODUCTION

Bladder cancer (BC) is the sixth leading cancer diagnosed in the United States and the fourth among males.¹ In 2022, there were an estimated 81,180 new cases in the United States and 17,100 deaths. More than 95% of BC is urothelial carcinoma and at diagnosis approximately 75% of cases are non–muscle invasive BC (NMIBC), with the remainder being muscle invasive BC (MIBC) or metastatic disease.^2,3 BC has one of the highest lifetime treatment costs per patient and represents a significant challenge in oncology given its wide range of disease risks, management options, and prognoses.⁴

NMIBC recurrence rates are 70% at 3 years, and up to 20% progress to MIBC, driving intensive surveillance schedules and high cost.⁴ Numerous guidelines have stratified patients by their risk of recurrence and progression based on clinical and pathologic factors.³ These risk groups are used as the basis for surveillance scheduling and decision making regarding treatment. Appropriate risk stratification, however, remains challenging because of heterogenous groupings and variable quality of surgical resections. Such issues present a particular issue to clinicians in counseling patients on their disease risks, selecting appropriate adjuvant intravesical therapies, and deciding when to pursue radical surgery. Therefore, tools that aide in cancer detection, risk stratification and prognostication, improvement of the quality of care, and in decision making are needed to improve the overall quality of NMIBC management.

In MIBC, prognosis depends on stage of disease at time of treatment. For years, the mainstay of management consisted of neoadjuvant, platinumbased chemotherapy followed by radical cystectomy (RC) with urinary diversion.⁵ This strategy itself represents a challenge to providers because of poor tolerability of platinum chemotherapies by a significant subset of patients with BC, unknown benefit in variant histology, and the morbidity associated with RC. To combat this, there has been an uptake in trimodal therapy for MIBC, which consists of radical transurethral resection of bladder tumor (TURBT) combined with chemotherapy and radiation.^5,6 Initially limited to small, solitary, fully resected lesions, the use of trimodal therapy has grown to capture a larger swath of patients who are poor surgical candidates. Ideal patient selection remains uncertain and represents a key unmet need in urology. Likewise, the selection of patients most likely to benefit from neoadjuvant chemotherapy, or selection between different agents as they become available, represents another challenge in MIBC. In metastatic disease, identification of patients at high risk of recurrence, early detection of metastasis, and selection of systemic treatment regimens and sequencing of these regimens remain challenging.

In recent years, artificial intelligence (AI) and machine learning (ML) have emerged as potentially powerful tools in clinical medicine, particularly in medical image analysis and genomics.^7,8 AI and ML are poised to play a major role in addressing the unmet needs in BC, including diagnostic cystoscopy, pathologic diagnosis, molecular biomarkers, risk stratification, treatment assessment, and outcome prediction. In this review, we discuss new and emerging applications of AI in BC (Fig. 1).

Fig. 1.

Application of artificial intelligence for bladder cancer diagnosis and outcome prediction. AI can be used for bladder cancer diagnosis and outcome prediction. Subapplications of bladder cancer diagnosis include tumor detection, staging, grading, and segmentation. For outcome prediction, prediction of recurrence, survival, and chemotherapy response have also been investigated using AI. Challenges to be addressed before integration into the clinical workflow include ensuring that AI models be generalizable, interpretable, and specific enough to not overburden the existing health care system with false-positive diagnoses. CT, computed tomography; H&E, hematoxylin-eosin.

Artificial Intelligence Background

There is growing interest in the development of AI as a tool to address some of the challenges in BC.⁹ Rooted in the disciplines of optimization, cognitive science, probability theory, and statistical learning theory, AI uses computer science principles and large datasets to solve problems.¹⁰ Subsumed by AI is the field of ML, a domain specifically focused on the development and optimization of intelligent, computer-run algorithms trained on large datasets that are designed to accomplish a wide variety of tasks.¹¹ ML itself subsumes deep learning, which is most relevant as a potential decision aid for the management of BC. Like ML, deep learning uses large datasets and computer-run algorithms to solve a wide variety of problems.¹² Deep learning differs from other subfields of ML in that it directly learns features from raw data. Although conventional ML methods require the development of feature extractors that convert raw data to representative embeddings, deep learning performs this representation learning step implicitly in its formulation.

At its core, deep learning is composed of a series of artificial neural networks (ANNs) that extract relevant features from data to generate a desired output.¹³ Many deep learning models are optimized using gradient descent, an iterative algorithm that uses the model’s prediction and data sample’s corresponding label to optimize its weights in a supervised setting. Deep learning models have been developed and applied to process and represent a wide variety of data modalities, including but not limited to tabular data, text, images, and videos.¹⁴ Of particular interest to BC is the development of deep learning algorithms for image processing called convolutional neural networks (CNNs).¹⁵ Designed to “see” semantic objects in images by representing complex, recognizable image features as compositions of representations of simpler features, CNNs have demonstrated success in numerous image processing tasks, such as image classification, object detection, instance segmentation, and semantic segmentation.^16,17

AI has the potential to be used in BC detection, treatment, and outcome prediction. Detection is further subdivided into imaging-based tumor identification and tissue/molecular biomarker-based assessment of cancer grade, stage, and risk stratification. Promising applications of AI for BC imaging include integration with cystoscopy, CT, MRI, and ultrasound to identify and localize bladder tumors. AI has also been used to assess tumor stage and grade using molecular biomarkers and cytology samples. For treatment and outcome prediction, AI has demonstrated potential for prediction of postoperative outcomes after RC, radiation therapy treatment planning, and recurrence likelihood prediction. However, such challenges as generalizability, interpretability, and overdiagnosis remain to be addressed. In the following sections, we review and highlight the current state for each of these applications.

ARTIFICIAL INTELLIGENCE FOR BLADDER CANCER DETECTION: IMAGING ANALYSIS

Cystoscopy

Initial detection of BC most commonly occurs with diagnostic cystoscopy during evaluation for hematuria. Several studies have investigated integration of deep learning models during cystoscopy (Table 1).^18-21 The availability of carefully annotated imaging datasets of pathologically confirmed bladder tumors (Fig. 2) is an important starting point for model training.¹⁸ In 2019, Shkolyar and colleagues¹⁸ reported CystoNet developed to detect bladder tumors during cystoscopy and TURBT. The model was trained on 611 manually annotated images containing pathologically confirmed BC, and then validated on a dataset of more than 7000 frames of confirmed BC derived from 44 tumors and 26 unique cases; and more than 50,000 frames of normal/benign urothelium from 31 cases. The overall per frame sensitivity and specificity were 90.9% and 98.6%, respectively.¹⁸ On a per-tumor basis, 39 out of 41 papillary tumors were detected and three out of three flat lesions were detected, yielding a final per-tumor sensitivity of 95.5%.¹⁸ More recently, Chang and colleagues²² implemented CystoNet in a real-time setting during TURBT and demonstrated initial feasibility of intraoperative integration.

Table 1.

Overview of studies using AI for bladder cancer detection

Study	Data (Validation)	Model	Performance	Comments
Shkolyar et al,18 2019	57 cases; 44 tumors; 59,515 frames (7542 tumor, 51,973 normal)	CNN with overlayed object detection RPN	90.9% per-frame sensitivity 95.5% per-tumor sensitivity 98.6% per-frame specificity	39 of 41 papillary tumor detected 3 of 3 flat tumors detected
Ikeda et al,19 2020	124 cases; 431 tumors; 431 frames (87 tumor, 335 normal)	Pretrained GoogleNet with fine-tuned cystoscopy dataset	89.7% per-frame sensitivity 94.0% per-frame specificity 0.98 AUC	Tumor frame distribution:  61.5% elevated  17.6% flat  20.9% mixed
Ali et al,20 2021	216 cases; 216 tumors; 216 BLC images (all cancer); 4 clinic centers	4 pretrained CNNs fine-tuned with BLC images:  VGG16  ResNet50  InceptionV3  MobileNetV2	Benign vs malignant:  95.8% per-frame sensitivity  87.84% per-frame specificity Invasiveness:  88.0% per-frame sensitivity  96.56% per-frame specificity	Performance obtained via L10PO-CV Best-performing models exhibited superior performance over 2 board-certified urologists
Wu et al,27 2022	1427 cases; 647 tumors; 5752 frames (647 tumor, 5105 normal), 4 clinical centers	Pretrained pyramid-scheme parsing network fine-tuned with cystoscopy dataset	95.0% per-frame sensitivity 98.7% per-frame specificity	Best-performing models exhibited superior performance over 9 urologists with varying degrees of experience

Abbreviations: AUC, area under the curve; BLC, blue-light cystoscopy; RPN, recurrent proposal network.

Fig. 2.

Representative sample of the types of bladder tumors encountered during TURBT overlayed with expert-verified annotations. Here one can observe the wide variety of tumor morphologies, locations, and sizes, along with various pathologic assessments (not pictured).

Ikeda and colleagues¹⁹ developed and applied a CNN to classify a cystoscopy dataset consisting of 2102 images. In this dataset, 1671 images contained normal tissue, and 431 images contained pathology-confirmed BC with a range of stages, grades, and morphologies. Here, a CNN pretrained on the ImageNet dataset²³ (GoogLeNet²⁴) was fine-tuned with their cystoscopy dataset. On the unseen test data, this transfer learning approach achieved an area under the receiver operating characteristic score of 0.98, a sensitivity of 89.7%, and a specificity of 94.0%. Similar studies have been performed that arrive at similar conclusions. Eminaga and colleagues²⁵ validated the performance of seven computer vision models with cystoscopy images originating from a digital atlas, and Yoo and colleagues²¹ demonstrated the benefit of using tumor color as an engineered feature to predict bladder tumor presence and grade.

Another study described the use of frame-based deep learning models to predict bladder tumor grade and stage using blue-light cystoscopy images sourced from multiple centers.²⁰ Blue-light cystoscopy is a widely used enhanced cystoscopy technology that involves intravesical instillation of an imaging agent, hexaminolevulinate, which causes cancerous urothelium to selectively fluoresce under blue light.²⁶ The study compared tumor staging, grading, and malignancy status using four CNNs with two clinicians. Although limited by small sample size of clinicians, the results suggest that deep learning has the potential to improve prediction of tumor stage and grade at the time of visual inspection. Findings from Wu and colleagues²⁷ support this claim whereby their model detected lesions with a much higher sensitivity than that of trainee, competent, and expert urologists. As an alternative to frame-based models described previously, video processing deep learning models that use sequential images for model training may offer improved performance in real-world settings.²⁸

Computed Tomography, MRI, and Ultrasound

CT, MRI, and ultrasound are widely used in urology and specifically in BC, ranging from initial diagnosis and staging to treatment response assessment and surveillance. AI may improve the performance of these imaging modalities, including automation of time-consuming tasks, tumor segmentation, classification, and treatment planning.^29,30 Furthermore, AI may enhance the identification of subtle imaging features and facilitate the surveillance of disease progression and treatment response.

CT scans are widely used for the initial evaluation and staging of BC. They may provide information about the tumor’s location, size, and regional lymph node involvement.^31,32 Contrast-enhanced CT scans are particularly useful in detecting metastasis. Traditional CT image analysis methods rely on manual interpretation, which is time-consuming and prone to human error. However, with the advent of deep learning, computer algorithms can be trained to accurately segment and classify lesions seen on CT images.

MRI provides detailed soft tissue contrast, making it particularly useful for assessing muscle involvement and staging.^33,34 MRI holds promise in differentiating between NMIBC and MIBC, and identifying the involvement of adjacent structures, such as the prostate or pelvic sidewall.³⁵ By combining various MRI sequences, including T2-weighted, diffusion-weighted, and dynamic contrast-enhanced imaging, radiologists can obtain comprehensive information about tumor characteristics. Deep learning algorithms applied to MRI data can automate tumor segmentation, aid in identifying suspicious regions, and provide quantitative analysis of tumor features, with such methods demonstrating promise for other organ systems, such as the brain.^36,37

Deep learning techniques, such as CNNs and recurrent neural networks, have shown potential in improving BC imaging.^31-34 By training these algorithms on large datasets of annotated images, complex patterns and features can be discerned. Deep learning models can accurately segment bladder tumors, classify them according to malignancy, and predict patient outcomes based on imaging features. These models can also be integrated into computer-aided diagnosis systems,³⁸ which assist radiologists in their interpretation of CT and MRI scans, reducing diagnostic errors and improving overall efficiency. Furthermore, deep learning algorithms can aid in the identification of subtle changes in tumor size, shape, or enhancement patterns over time, enabling better monitoring of treatment response and early detection of disease recurrence, with successful findings demonstrated for other organ systems.³⁹

Ultrasound is a noninvasive and safe imaging modality that uses high-frequency sound waves to generate images of the body’s internal structures.^40,41 Ultrasound can be used repeatedly without significant risks^42,43 and has been investigated for BC screening and surveillance.^44,45 Transabdominal ultrasound may be useful for tumor detection and assessment of size and location in some instances.^46,47 In contrast to cystoscopy, it can evaluate surrounding structures to check for tumor spread. Although not routinely used, transrectal ultrasound can provide detailed images of the bladder lumen, especially for tumors near the bladder neck or urethra.^48,49 Contrast-enhanced ultrasound improves visualization of vascularized bladder tumors.^50,51 Recent advances in three-dimensional and four-dimensional ultrasound offer volumetric images that enhance visualization and assessment of tumor characteristics.⁵² Application of deep learning to enhance ultrasound imaging of BC is at an early stage. Deep neural networks have been trained on annotated ultrasound images to recognize tumor features and differentiate between benign and malignant lesions. Recent work has shown that deep learning is promising in automating tumor segmentation, identifying suspicious regions, and assisting in treatment planning.⁵²

ARTIFICIAL INTELLIGENCE FOR BLADDER CANCER DIAGNOSIS: GRADING, STAGING, AND MOLECULAR BIOMARKERS

Cytology, Pathology, and Risk Stratification

Urine cytology plays an important role in the diagnosis and surveillance of BC given its noninvasive nature. Numerous studies have used ML to accurately identify and classify cancerous cells, achieving high sensitivity and specificity.^53-56 This AI-based cytology analysis holds potential for enhancing the speed and accuracy of BC diagnosis. Pathologic examination of resected tumor tissue is the recognized standard in cancer diagnosis and risk stratification.^57-59 AI-powered pathology tools offer promise in improving diagnostic accuracy and aiding in risk stratification for patients with BC, which is essential for guiding treatment decisions and prognostication.^60-62 Numerous groups have developed deep learning models that predict disease progression in patients with BC based on clinical and pathologic features.^60,61,63,64 These AI-based risk stratification models may enable clinicians to identify high-risk patients who require more aggressive treatment or modified surveillance schedules. The integration of AI with cytology, pathology, and risk stratification processes offers promising avenues for enhancing BC management and improving patient outcomes.

Molecular Biomarkers and Gene Signatures

There has been significant effort aimed at identifying molecular biomarkers and gene signatures that can aid in the diagnosis, prognosis, and treatment of BC.^65-69 One of the most extensively studied molecular biomarkers in BC is the fibroblast growth factor receptor 3 (FGFR3) gene.^70,71 Mutations in the FGFR3 gene are prevalent in NMIBC and have been associated with a favorable prognosis. Another important biomarker is the telomerase reverse transcriptase (TERT) gene promoter mutation.⁷² TERT promoter mutations are found in a significant proportion of BC cases and have been associated with aggressive tumor behavior and poor prognosis. Gene expression profiling is a powerful tool for identifying gene signatures associated with BC. Several studies have used high-throughput gene expression technologies, such as microarrays and RNA sequencing, to identify gene expression patterns that can predict disease outcomes and treatment responses.^67,68,73 One of the well-known gene signatures in BC is the Lund Taxonomy,^74,75 which classifies tumors into different molecular subtypes based on gene expression profiles. Gene signatures derived from the analysis of specific gene sets or pathways have also been investigated in BC. For example, the P53 pathway signature is associated with poor prognosis in BC.^75,76 Other gene signatures, such as those related to cell cycle regulation, DNA repair mechanisms, and immune response pathways, have shown potential as prognostic and predictive markers.

Deep learning–based gene signatures are increasingly being used for the prediction and prognosis of BC.^77,78 This approach leverages AI to analyze large-scale genomic data, identifying unique gene expression patterns associated with BC. These gene signatures, often composed of multiple genes, can serve as biomarkers for the disease, aiding in early detection, prediction of disease progression, and personalization of treatment strategies. The use of deep learning algorithms allows for the processing of complex, high-dimensional data, uncovering intricate relationships between genes and disease phenotypes that may not be readily apparent through traditional analysis methods.

ARTIFICIAL INTELLIGENCE FOR BLADDER CANCER TREATMENT AND OUTCOME PREDICTION

Predicting Postoperative Morbidity and Mortality

Although many cases of BC are managed via endoscopic resection with or without intravesical therapy, extirpative surgery remains the intervention of choice in patients with NMIBC at high risk for progression and in MIBC.^3,79,80 However, RC mortality rates are 1% to 3% at large centers.⁸¹ More than 60% of patients experience at least one complication within 90 days of cystectomy, with 13% being grade 3 or higher.⁸² Perioperative comorbidity indices, such as the adjusted Charlson Comorbidity Index (aCCI), American Society of Anesthesiology classification system (ASA), and National Surgical Quality Improvement Program risk calculator, are poor at predicting postoperative morbidity after RC.^83,84 Several recent studies have proposed supervised ML models that use preoperative variables to improve RC risk assessment and optimize patient selection for RC.

Klén and colleagues⁸⁵ trained a logistic regression (LR) classifier with Least Absolute Shrinkage and Selection Operator (LASSO) regularization to predict early death after RC. LASSO is a powerful and well-established method of pruning weakly predictive variables out of a linear model. This allowed the investigators to use a variety of preoperative factors (comorbidities and their indices, patient demographics, laboratory values) with a lower risk of overfitting their model. They trained their model on 733 patients from 16 different hospitals and reported an area under the curve (AUC) of 0.73 in an independent test set of 366 patients. Congestive heart failure, ASA class, and chronic pulmonary disease were the strongest predictors of postoperative mortalities.⁸⁵

This approach, the well-established LR classifier with LASSO regularization, remains one of the most effective models in comparison with more complex ML algorithms. In a larger comparative study of 7557 patients from the National Surgical Quality Improvement Program risk calculator database who underwent RC, Taylor and colleagues⁸⁶ developed several supervised ML models for predicting adverse events, extended hospital course, and discharge to higher level of care. They carried out model training and 10-fold cross-validation using 80% of the dataset to show excellent model stability and minimal overfitting with a similarly expansive set of preoperative inputs. Selected algorithms included a generalized additive model, a neural network, a random forest model, and LASSO. LASSO outperformed all other models in the reserved test set. AUC was 0.63 for predicting adverse events, 0.68 for extended hospital stay, and 0.75 for discharge to a higher level of care.

Supervised ML outperforms individual comorbidity indices in predicting morbidity after RC. A 2022 study of 392 patients directly compared the predictive power of several supervised ML models with aCCI, ASA, and Gagne’s combined comorbidity Index (GCI). AUC for comorbidity indices were 0.6 (aCCI), 0.63 (ASA), and 0.58 (GCI). In contrast, the AUC for LR was 0.763, second only to the Gaussian Naive Bayes model, which performed with an AUC of 0.794. Both the neural network and random forest also outperformed comorbidity indices with AUCs of 0.741 and 0.748, respectively.⁸⁷

A handful of studies explore ML models with more novel inputs. Ying and colleagues⁸⁸ deployed a CNN to automatically segment and compute skeletal muscle volume on preoperative CT before RC. Cox regression analysis in their cohort of 299 patients showed independent association with overall survival (hazard ratio, 1.62; 95% confidence interval, 1.07–2.44; P = .022). Schuettfort and colleagues⁸⁹ built a LASSO model to predict various oncologic outcomes based on systemic inflammatory response biomarkers in 4199 patients. AUC was 0.673 for predicting lymph node involvement, 0.73 for predicting greater than or equal to pT3 stage, and 0.658 for upstaging to muscle-invasive cancer on final pathology.

Predicting Postcystectomy Survival and Cancer Control

Long-term cancer control after RC is poor. Even with negative margins, 5-year overall survival rates without neoadjuvant chemotherapy are approximately 47%.⁹⁰ Accurate prognostication can help determine surveillance schedules and candidacy for adjuvant systemic therapy after primary resection. Postoperative cancer control is currently prognosticated primarily by pathologic staging.^3,79,80 A recent retrospective multicenter study of 9000 patients reported a concordance index of 0.68 when standard American Joint Committee on Cancer TNM staging was used to predict 5-year progression-free probability after RC. The authors were able to improve the index to 0.75 using a Cox proportional hazards model to build a nomogram, although this was not based on an external test set.⁹¹ Several studies have sought to improve prognostication with ML techniques.

Early efforts began in the early 2000s to train ANNs to predict RC outcomes using standard perioperative clinical data including pathologic stage, other adverse surgical pathologic findings (eg, lymphatic invasion), age, gender, and surgical approach. The results were equivocal. In 2006 Bassi and colleagues⁹² developed an ANN to predict 5-year overall survival after RC and compared its predictive accuracy with an LR classifier developed from the same population (n = 369). The simple regression model performed just as well as the ANN (75.9% vs 76.4% concordance index). In 2013 Buchner and colleagues⁹³ used a population of 2111 patients to train an ANN to predict recurrence, cancer-specific mortality, and all-caused death. A Cox regression model was developed from the same population for comparison. The ANN was 74% accurate at predicting recurrence, 69% for cancer-specific mortality, and 69% for all-caused death. The Cox regression model was only less accurate by 4.7% (P < .001) at predicting cancer-specific mortality and 3.5% (P = .007) for predicting all-caused death. There was no difference in accuracy at predicting recurrence. Results were also reported without the use of a test set to demonstrate external validity.⁹³

Later studies used various other supervised ML models, also with mixed results. In 2009 Catto and colleagues⁹⁴ used 609 patients with localized BC to train a neuro-fuzzy model to predict recurrence after RC based on standard clinicopathologic factors. They applied two published nomograms to predict recurrence in their cohort as a comparator. The model was 84% accurate, compared with 72% and 74% with the two nomograms. Although the authors did report performance using a held-out test set, the nomograms were developed from a different study population entirely, thus introducing selection bias and decreasing external validity in favor of the neuro-fuzzy model.⁹⁴ In 2015 Wang and colleagues⁹⁵ used data from 117 patients to train seven supervised ML methods to predict 5-year survival after RC. The accuracy of these ML models ranged from 70% to 80%, compared with 63.3% using a published nomogram. However, results were once again reported based on the same study cohort used to train the ML algorithms but not the nomogram. Although the authors did use 10-fold cross-validation to report misclassification probability (standard deviation for mean accuracy among folds approached ±8%), no external test set was used, which all but guarantees that the nomogram will perform worse.⁹⁵

In 2019, Hasnain and colleagues⁹⁶ published the largest study to date applying supervised ML algorithms to predict recurrence and survival after RC (n = 3499). This study notably used a more expansive set of input variables including medical comorbidities, neoadjuvant chemotherapy, surgical metrics and methods, and various other perioperative variables beyond pathologic stage and other standard clinicopathologic variables. ML models were stacked into “meta-classifiers” using various summation methods, such as “mixture-of-experts” and hard-voting. As a comparator, recurrence and survival were also predicted using pT stage in a univariate model. The meta-classifier was 70% sensitive and 70.2% specific in predicting 5-year recurrence. In comparison, isolated pT stage was 74.4% sensitive and 61.1% specific. Results were reported using nested cross-validation (10 outer folds, 5 inner folds), but not with a held-out test set.⁹⁶

Sonpavde and colleagues⁹⁷ recently published the results of a study that trained an ANN with whole exome sequencing of tumor and normal tissue using RNA-Seq. The classifier predicted recurrence with an accuracy of 94% using test sets. The study is small (n = 117), but the preliminary results are promising.⁹⁷

Limitations of the Existing Body of Work

Compared with standard regression models, such as LR classifiers, ANNs decrease the transparency of analysis by incorporating one or more hidden layers of nonlinear activation functions. Other nonlinear supervised ML models have a similar black box effect. This forfeits the benefit of an exploratory analysis, which would otherwise provide odds ratios for isolated variables to help inform a clinician’s decision-making process. Granted, the strategy of AI and ML is to sacrifice a mechanistic understanding of the prediction problem in favor of an effective black box solution. This trade-off is reasonable, provided that a study formally states its goal to be predictive rather than exploratory, and a study takes strict measures to validate and test the black box. ANNs and more complex ML models are at higher risk of overfitting the data, especially without rigorous methodology to report misclassification probability and appropriate performance metrics.⁹⁸

In this respect, limitations remain in the existing body of literature in regards to methodology. Many studies either lacked a held-out test set or reported performance using a test set from the same population that trained the ML algorithm but not the comparator. Existing studies in this space also tend to report accuracy as the primary performance metric, which is less representative of true performance when there is class imbalance. In the absence of neoadjuvant chemotherapy, about 80.9% of patients have disease recurrence within 2 years of RC.⁹⁰ Additional statistics, such as the AUC and the F1 score, are often better-suited but not always used.⁹⁹ The choice of model is often also suboptimal. More complex supervised ML models are designed to elucidate a network of nonlinear, nonintuitive relationships between a large array of inputs. They provide minimal benefit when given a simple set of a few clinicopathologic variables.⁹⁸ Indeed when proper methods were deployed, they were often outperformed by simple logistic-regression-based models, such as LASSO.^86,87 The latest studies have begun demonstrating appropriate model selection and proper methodology, but further prospective study is needed to demonstrate clinical value.^87,88,90

A final major obstacle to deploying effective AI solutions in this space is the issue of “decaying relevance.” Training ML models requires large swaths of clinical data to iteratively home in on a solution. It takes many years to generate the requisite volume of survival and recurrence data after RC. Chen and colleagues¹⁰⁰ demonstrated that predicting future inpatient orders is vastly more accurate with models trained on just 1 month of recent data than with 12 months of older but higher volume data. This is because of the speed at which clinical practice patterns evolve. For example, many patients with MIBC today receive neoadjuvant chemotherapy or radiotherapy, whereas most studies to date exclude patients who received any presurgical treatment. This severely limits the relevance of these AI tools on today’s cystectomy candidates.

Application of Artificial Intelligence for Radiation-Based Treatment of Bladder Cancer

Radiation therapy is an important treatment modality in patients with MIBC.^101-103 Radiation therapy is used as a primary treatment of BC in cases where surgical intervention (ie, RC) is not possible or preferred by the patient,^104,105 or as palliative treatment to alleviate symptoms and improve the quality of life for patients with advanced or metastatic disease.

One of the challenges of radiation therapy for BC is minimizing the radiation dose to healthy surrounding organs, such as the bowel and rectum. Modern external beam radiation therapy,¹⁰⁶ including intensity-modulated radiation therapy, and image-guided radiation therapy allow for better sparing of healthy tissues, reducing the risk of complications and improving patients’ overall tolerance to treatment.

Deep learning–based radiation therapy is emerging as a promising approach for the treatment of BC. This technique uses AI to optimize the planning and delivery of radiation therapy, aiming to maximize the dose to the tumor while minimizing exposure to surrounding healthy tissues.^107,108 Deep learning models can be trained to automatically segment the bladder and other critical structures in imaging data, a task that is traditionally time-consuming and prone to interobserver variability.^109,110 These models can also predict the movement and deformation of the bladder during treatment, allowing for more precise targeting of the radiation dose. Although at an early stage, deep learning has the potential to improve more precise delivery of radiation therapy for BC.

SUMMARY

AI holds strong potential to guide decision-making for BC diagnosis and treatment and to integrate such models into the current clinical workflow. This review summarizes the emerging AI applications for BC, including diagnostic cystoscopy, radiologic imaging, pathologic diagnosis, molecular biomarkers, risk stratification, treatment assessment, and outcome prediction.

Moving forward, several areas of development and integration of AI systems are needed: (1)curation of large, multicenter, multidimensional datasets for training and validation; (2) standardization of reporting and data management to facilitate communication; and (3) design of prospective clinical trials to validate the utility of these models in key timepoints of cancer management from screening, diagnosis, treatment, and surveillance. The myriad of data modalities collected in the BC clinical workflow, from clinical history, cystoscopy, cross-sectional imaging, laboratory data, and histopathology, can be effectively used in new multimodal deep learning models. The recent advent of artificial general intelligence and large language models, such as ChatGPT have sparked their application for writing and reasoning tasks in clinical and surgical specialites.¹¹¹

Successful execution of the previously mentioned aims will have wide-reaching implications for patients, providers, and health care systems. The integration of AI into the current BC management clinical workflow will serve as a personalized decision support system for patients and clinicians throughout its detection, treatment, and downstream management, and will ultimately result in improved patient outcomes and quality of life.

KEY POINTS.

• Bladder cancer is a costly, highly recurrent disease that requires lifetime surveillance after initial diagnosis, and its early detection is critical to optimize patients’ expected survival and quality of life.

• Several studies have investigated the potential for AI to improve the effectiveness of the bladder cancer clinical workflow in its detection and treatment phases.

• Significant effort remains to validate the effectiveness of AI for bladder cancer clinical decision support.