Artificial Intelligence: The Latest Advances in the Diagnosis of Bladder Cancer
Article information
Abstract
Bladder cancer remains a significant health challenge. Early and accurate diagnoses are crucial for effective treatment and improved patient outcomes. In recent years, artificial intelligence (AI) has emerged as a powerful tool in the medical field, showing great promise in advancing the bladder cancer diagnosis. This review explores the current state and potential of AI technologies, including machine learning algorithms, deep learning networks, and computer vision, in enhancing the diagnostic process for bladder cancer. AI systems can analyze vast amounts of data from various sources, such as medical imaging, genomic data, and electronic health records, enabling the identification of subtle patterns and biomarkers that may indicate the presence of bladder cancer. These systems have demonstrated high accuracy in detecting cancerous lesions in imaging modalities such as cystoscopy, ultrasonography, and computed tomography scans, often surpassing human performance. Moreover, AI-driven diagnostic tools can assist in risk stratification, predicting disease progression, and personalizing treatment plans, thereby contributing to more targeted and effective therapies.
INTRODUCTION
Bladder cancer is one of the most common malignancies, with significant morbidity and mortality rates worldwide [1]. Traditional diagnostic methods, such as cystoscopy, urine cytology, and imaging techniques, are limited by their sensitivity, specificity, and invasiveness. According to Russell and Norvig, “AI is a program that acts like a human (Turing test), thinks like a human (human-like patterns of thinking steps), and acts and thinks rationally (logically, correctly).” AI has a long history, originating in 1943 with McCulloch and Pitts, who simplified the mathematical model of neurons realize all propositional logic primitives. The term “AI” was first coined by John McCarthy in 1956 [2]. The advent of artificial intelligence (AI) has the potential to revolutionize the field of bladder cancer diagnosis, offering more accurate, efficient, and noninvasive methods.
Cystoscopy is the gold standard for bladder cancer diagnosis. It involves the direct visualization of the bladder mucosa using a cystoscope, allowing for biopsy and histopathological examination. Despite its accuracy, cystoscopy is invasive, uncomfortable for patients, and associated with complications such as infection and bleeding. Urine cytology involves the microscopic examination of urine samples for cancerous cells. While it is noninvasive, its sensitivity is relatively low, particularly for low-grade tumors. It is often used in conjunction with other diagnostic methods to increase accuracy. Imaging techniques such as intravenous urography, ultrasound, computed tomography (CT), and magnetic resonance imaging (MRI) provide valuable information about the presence and extent of bladder tumors. However, these techniques have limitations in detecting small or flat lesions and often require contrast agents, which may not be suitable for all patients. AI applications for the diagnosis of bladder cancer integrate imaging with bladder segmentation, tumor detection on cystoscopy, tumor staging, and tumor grading [3]. AI technologies, including machine learning (ML) algorithms, deep learning (DL) networks, and computer vision, are revolutionizing diagnostic processes by analyzing vast amounts of data from medical imaging, genomic data, and electronic health records (EHRs) [3-5].
METHODOLOGY
We conducted a comprehensive review to enhance our understanding of AI applications, promote familiarity with these technologies, and identify future research opportunities in the advanced diagnosis of bladder cancer. Additionally, we aimed to assess the feasibility of integrating these applications into clinical practice. To achieve this, we searched the PubMed, Web of Science, Embase, and MEDLINE databases for original articles published in the last 15 years, using keywords such as “artificial intelligence,” machine learning, “diagnosis,” deep learning and “bladder cancer.” We included studies up to May 2024 that investigated all diagnostic methods or modalities for bladder cancer, whether for initial diagnosis, recurrence, or progression. A total of 50 articles were searched, and 26 were found to be relevant and included in this review. Searches were conducted in PubMed and other databases indexed with MeSH (medical subject headings) terms, following the guidelines from the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses). We excluded irrelevant topics, studies without full-text access, and those with inadequate data. Datasets were manually reviewed to detect and eliminate duplicate entries or data points (Fig. 1).
PRISMA (Preferred Reporting Items for Systematic Reviews and Meta- Analyses) study selection flowchart.
RESULTS
The primary data sources for AI models in bladder cancer diagnosis include medical imaging (CT, MRI, ultrasound), histopathological images EHRs, and genomic data. These images undergo preprocessing steps such as normalization, noise reduction, and contrast enhancement. Additionally, techniques including resizing, rotation, and cropping are used to augment the dataset. EHR data are normalized, anonymized, and subjected to feature extraction to ensure consistency and privacy. Feature extraction is performed using convolutional neural networks (CNNs), which can automatically detect patterns and anomalies in imaging data. Similarly, natural language processing and bioinformatics tools are utilized to identify relevant biomarkers and clinical parameters. Labeled datasets, which indicate the presence or absence of bladder cancer, are employed to train ML models such as support vector machines (SVMs), random forests, and neural networks. DL models, particularly CNNs, are trained on extensive datasets of medical images to learn hierarchical features for tumor detection and classification [6]. Techniques like k-fold cross-validation are used to validate the model’s performance and prevent overfitting. Metrics such as accuracy, sensitivity, specificity, precision, recall, and the area under the receiver operating characteristic curve (AUC) are calculated to assess the model. AI models are integrated into clinical decision support systems (CDSS) to aid radiologists and pathologists in diagnosing bladder cancer. Intuitive interfaces are developed to display AI-generated insights to clinicians, highlighting areas of concern in imaging studies and providing risk scores based on EHR data. Recent advancements in AI have shown significant promise in improving the diagnosis of bladder cancer, a condition that continues to pose a substantial health challenge due to its high morbidity and mortality rates.
AI-driven systems have shown remarkable accuracy in detecting cancerous lesions across various imaging modalities, including cystoscopy, ultrasound, and CT scans, often exceeding the capabilities of human experts. The use of CNNs for automated cystoscopy analysis enables the identification of malignant lesions with high sensitivity and specificity, thereby reducing the subjectivity and variability associated with manual interpretation. Additionally, radiomic analyses of CT and MRI scans enhance diagnostic capabilities by extracting quantitative features that are invisible to the human eye, which aids in distinguishing between benign and malignant lesions.
Moreover, urine cytology has been improved by utilizing ML algorithms to analyze cell morphological characteristics, thus improving the sensitivity and specificity in detecting malignant cells. Furthermore, the integration of AI with genomic and proteomic data from urine samples has facilitated the creation of highly accurate, noninvasive diagnostic models [7].
Despite these promising advancements, challenges such as ensuring data quality and standardization, addressing ethical and legal considerations, and integrating AI tools seamlessly into clinical workflows remain. Improving the interpretability and explainability of AI models is crucial for gaining clinicians’ trust and facilitating their adoption in clinical practice.
Examples of AI basics are summarized in Fig. 2.
Examples of artificial intelligence basics. (A) Machine learning versus deep learning. (B, C) Neural network (basic intuition). (D) Deep learning (basics). MNIST, Modified National Institute of Standards and Technology.
DISCUSSION
AI, particularly ML and DL algorithms, has shown promise in improving the accuracy and efficiency of bladder cancer diagnosis. These technologies can analyze vast amounts of data, recognize patterns, and make predictions with high precision. Nonetheless, implementing AI models in clinical settings presents several challenges, including data standardization, ethical concerns, and regulatory compliance.
1. Data Standardization
AI models require extensive datasets for training and enhancing their performance, however, standardizing medical data presents a significant challenge.
1) Diverse data sources
Bladder cancer diagnosis involves various data sources, including medical imaging, histopathology, and clinical records. Each of these sources comes from different equipment manufacturers, hospitals, or even countries, leading to variability in data formats, resolutions, and quality. Imaging modalities like CT scans, MRI, and cystoscopy can vary in terms of acquisition protocols, making it difficult for AI models to generalize across different data. Pathology data from tissue biopsies may differ in staining techniques or microscopic resolution, adding complexity to AI applications.
2) Data annotation
High-quality labeled data is crucial for training AI models; however, annotating medical images and clinical data is both time-consuming and costly. This process can result in inconsistencies in labeling, particularly when different experts are involved in providing annotations, which may adversely affect the model’s performance.
3) Interoperability and integration
For AI systems to be effectively adopted in clinical settings, they must integrate seamlessly with existing EHR systems. However, EHR systems frequently lack standardized formats and coding languages, such as HL7 or FHIR, which complicates integration efforts. Without standardized data structures, AI models may be unable to access the comprehensive patient data required for accurate diagnoses.
2. Ethical Issues
AI introduces a range of ethical challenges, especially in the diagnosis of bladder cancer, where the safety and outcomes of patients are of utmost importance.
1) Bias in AI models
AI systems can inherit biases from the training data, a particularly concerning issue in healthcare where biased models can lead to unequal outcomes across different populations. For instance, if the training data primarily consists of patients from certain ethnic backgrounds or geographic regions, the AI model might not perform effectively for underrepresented groups. This disparity could lead to delayed or inaccurate diagnoses for minority populations, thereby worsening healthcare disparities.
2) Transparency and explainability
Many AI models, particularly those based on DL, function as black boxes, offering little transparency about how they formulate their decisions. In the context of bladder cancer diagnosis, where critical life-altering decisions are made, it is essential for both clinicians and patients to comprehend the reasoning behind the AI’s diagnoses. The absence of explainability raises ethical issues, as clinicians might be reluctant to depend on a system whose workings they do not fully understand. Explainable AI (XAI) seeks to remedy this issue; however, it remains an emerging field, particularly in intricate areas such as medical imaging.
3) Patient privacy
AI models frequently necessitate access to extensive patient data, which raises concerns about data privacy and confidentiality. In the context of bladder cancer diagnosis, there is a risk that sensitive information from patient medical histories, genetic profiles, or pathology reports could be exposed if not adequately protected. It is crucial to ensure compliance with privacy regulations such as the Health Insurance Portability and Accountability Act in the United States or General Data Protection Regulation in Europe. However, AI systems introduce additional complexities in ensuring that data sharing and processing adhere to these standards.
3. Regulatory Compliance
AI systems in healthcare, particularly those used for diagnosis, are subject to stringent regulatory oversight. Nevertheless, the regulatory landscape continues to evolve.
1) Approval pathways
For AI models to be utilized in clinical settings, they must receive approval from regulatory bodies such as the Food and Drug Administration (FDA) in the United States or the European Medicines Agency in Europe. These agencies are currently adjusting their approval processes to accommodate AI technologies. A significant challenge is deciding whether AI models should be classified as medical devices or decision support tools. Additionally, continuous learning systems, where AI models evolve as they process more data, pose further complications for regulatory frameworks because they are not static like traditional medical devices.
2) Clinical validation
AI models need extensive clinical validation to ensure their safety and efficacy prior to deployment. This process generally includes conducting prospective trials and validating the AI’s performance in real-world settings, which can be both costly and time-consuming. Furthermore, there is a risk that AI models might not perform as effectively in real-world clinical settings as they do in research environments. This discrepancy can be attributed to factors such as data variability, environmental differences, or issues with workflow integration.
3) Liability and accountability
In instances where AI-based diagnoses result in incorrect or delayed treatment, determining liability can be complex. It raises the question: should responsibility be attributed to the developers of the AI system, the healthcare provider, or the hospital? Establishing clear guidelines for accountability is crucial, yet these guidelines remain poorly defined.
4. Addressing the Challenges
1) Improving data standardization
Efforts are underway to develop more uniform data standards across institutions and countries. Initiatives such as the International Consortium for Health Outcomes Measurement are focused on creating standardized outcome measures, including those for bladder cancer. Additionally, collaborative data-sharing platforms across hospitals and research institutions can help reduce variability in datasets.
2) Mitigating bias
Addressing bias in AI models necessitates a deliberate effort to gather more diverse data that accurately represents a wide range of patient populations. Regularly auditing AI models for bias and implementing interventions like bias-correction algorithms are essential steps to help mitigate these effects.
3) Ensuring explainability
Ongoing research into interpretable AI models is essential for enhancing the transparency and acceptability of AI systems in clinical practice. Employing AI models that utilize visual heatmaps to emphasize areas of interest in medical images can elucidate the system’s decision-making process, thereby building trust among clinicians.
4) Navigating regulatory frameworks
AI developers must collaborate with regulatory bodies early in the design phase to ensure that their models comply with the necessary requirements. Another avenue for ensuring compliance as the model evolves is postmarket surveillance of AI systems, which involves continuous monitoring after deployment.
Below are some of the key areas where AI is making significant contributions. AI algorithms have been developed to assist in the analysis of cystoscopy images. These systems employ DL techniques, such as CNNs, to identify malignant lesions with high accuracy. For instance, studies have shown that AI can detect bladder tumors in cystoscopy images with sensitivity and specificity comparable to those of experienced urologists. Eminaga et al. [8] used CNN models to detect cancerous features in cystoscopy images for diagnostic purposes, with the Xception model demonstrating superior performance, achieving an F1 score of 99.52%. Lorencin et al. [9] implemented an artificial neural network to analyze frames from 1997 bladder cancer images and 986 images of benign or normal mucosa, achieving a remarkable AUC of 0.99. The same team [10] used a different CNN algorithm to differentiate between benign and malignant lesions in 2,983 cystoscopy images, also achieving an AUC of 0.99. Ikeda et al. [11] developed a CNN algorithm that demonstrated high sensitivity and specificity (89.7% and 94%, respectively) for cancer detection in 2102 cystoscopy images. Yang et al. [12] evaluated various CNNs (LeNet, AlexNet, and GoogLeNet) using the EasyDL platform to identify bladder cancer images, with EasyDL achieving the highest accuracy at 96.9%. Du et al. [13] applied a CNN-DL algorithm to recognize bladder cancer in images from 175 patients, with EasyDL providing the best accuracy at 96.9%.
Automated analysis minimizes the subjectivity and variability inherent in manual interpretation, potentially improving diagnostic accuracy and alleviating the burden on healthcare professionals.
Radiomics involves the extraction of quantitative features from medical images, which are subsequently analyzed using ML algorithms to detect patterns linked to bladder cancer. These features, frequently imperceptible to the human eye, offer crucial diagnostic and prognostic insights. For instance, radiomic analysis of CT and MRI images has proven effective in distinguishing between benign and malignant bladder lesions, predicting tumor grade, and evaluating treatment responses.
AI has been applied to enhance the sensitivity and specificity of urine cytology. ML algorithms are capable of analyzing the morphological characteristics of cells in urine samples, effectively distinguishing between benign and malignant cells with high accuracy. These systems support cytopathologists in more reliably identifying cancerous cells thereby reducing the rates of false negatives and false positives. Nojima et al. [14] employed a 16-layer Visual Geometry Group (16VGG) CNN to assess the effectiveness of urinary cytology in detecting malignant or high-grade lesions. The 16VGG CNN exhibited outstanding performance in differentiating cancerous tissue from benign tissue (AUC, 0.9890; F1 score, 0.9002), as well as in identifying invasive bladder cancer (AUC, 0.8628; F1 score, 0.8239) and high-grade bladder cancer (AUC, 0.8661; F1 score, 0.8218). Awan et al. [15] developed a method for the automatic identification of atypical and neoplastic cells, with the Xception model achieving the highest performance in their validation set (AUC, 0.99). Vaickus et al. [16] explored the potential of a hybrid DL and morphometric model to automate the Paris System for urine cytology. They analyzed whole slide images from 51 negative, 60 atypical, 52 suspicious, and 54 positive cases, achieving a 95% accuracy rate in detecting cell types and atypia, which could facilitate the automation of the Paris System. Sanghvi et al. [17] utilized a CNN algorithm to diagnose bladder cancer using cytology images from 2,405 Thin Prep glass slides and validated it on a separate dataset, achieving an AUC of 0.88 (95% confidence interval, 0.83–0.93), with a sensitivity of 79.5% and a specificity of 84.5% for high-grade urothelial carcinoma.
The integration of AI with genomic and proteomic data from urine samples has opened new avenues for bladder cancer diagnosis. AI algorithms can analyze complex datasets to identify biomarker signatures associated with bladder cancer. For example, ML models have been used to predict bladder cancer based on the expression levels of specific genes and proteins in urine, offering a noninvasive and highly accurate diagnostic tool. Additionally, some studies have investigated the role of urine metabolites in detecting bladder cancer, as these metabolites come into direct contact with the urine. Shao et al. [18] analyzed 87 bladder cancer samples and 65 control samples, identifying imidazoleacetic acid as a potential biomarker for bladder cancer. Using an ML model and a decision tree, they achieved an accuracy of 76.60%, with a sensitivity of 71.88% and a specificity of 86.67%. Kouznetsova et al. [19] focused on urine metabolites to classify bladder cancer through an ML model. They discovered that D-glucose, an early-stage bladder cancer biomarker, could influence neoplastic genes such as AKT, EGFR, and MAPK3.
AI is also utilized to develop predictive models for bladder cancer risk stratification. These models analyze patient data, including demographics, clinical history, and molecular profiles, to predict the likelihood of developing bladder cancer or experiencing disease recurrence. Such predictive analytics can guide personalized screening and surveillance strategies, enhancing early detection and patient outcomes. Garapati et al. [20] explored the potential of an automated ML technique to screen 84 bladder cancer tumors using CT urography. They classified the tumors as either T2 or higher or below T2, finding that using morphological and texture features alone or combined yielded similar performance. Xu et al. [21] focused on the preoperative staging of bladder cancer into non–muscle-invasive or muscle-invasive types. They used multiparametric MRI with 1,104 radiomic features from 54 patients to differentiate these stages. The model, developed using SVM-RFE and the synthetic minority oversampling technique, analyzed 19 features from T2-weighted and diffusion-weighted imaging sequences. This method achieved an AUC of 0.9857 for muscle invasion discrimination, surpassing expert diagnostic accuracy with a rate of 96.30%. Yin et al. [22] aimed to differentiate Ta from T1 bladder cancer using hematoxylin and eosin-stained images from 1,177 tissues. They employed an ML-CNN model along with image processing software like Image J and Cell Profiler, achieving accuracies between 91% and 96%.
Several AI algorithms have been developed and tested for diagnosing bladder cancer, each utilizing different techniques to improve accuracy and performance across various diagnostic contexts. Tables 1-5 present a direct comparison of the algorithms discussed.
Cystoscopy image analysis with convolutional neural networks (CNNs)
CNN and ANN models
Urine cytology and histopathology image analysis
Radiomics in bladder cancer diagnosis
Artificial intelligence in bladder cancer staging and risk stratification
5. Comparative Evaluation of Performance and Effectiveness
CNN-based algorithms such as Xception and EasyDL, consistently achieve high sensitivity, specificity, and accuracy in image-based bladder cancer diagnosis, often surpassing 95% in performance metrics. These algorithms are particularly effective in tasks like cystoscopy image analysis, where their accuracy matches or exceeds that of expert clinicians. Radiomics approaches that utilize CT or MRI scans excel in detailed tumor characterization and staging. When combined with ML techniques such as SVM or decision trees, these approaches achieve AUCs close to 1.0 for tumor staging. Urine cytology analysis models, including 16VGG and Xception, prove to be highly valuable for noninvasive cancer detection. They demonstrate high AUC scores and performance metrics in distinguishing malignant from benign cells. The integration of multiomics data using AI and ML, especially in analyzing genomic and proteomic data, is emerging as a promising field for noninvasive diagnostics. However, it currently falls short of the accuracy achieved by imaging-based techniques.
The risk of bias in individual studies is summarized in Table 6.
Risk of bias in individual studies
6. Future Potential and Advancements
1) Multimodal AI systems
Integrating diverse data types, such as imaging, genomics, proteomics, and clinical data, into a single AI model could enhance diagnostic accuracy and provide comprehensive insights into patient health. Combining imaging data with molecular signatures from urine or tissue samples could lead to more personalized diagnostic and prognostic tools.
2) Continuous learning AI models
AI models that continuously improve by learning from new data, such as those using reinforcement learning or self-supervised learning, have the potential to enhance real-time diagnostic systems in dynamic clinical environments. This approach could also help address challenges associated with model degradation over time.
3) AI in real-time CDSS
The integration of AI with CDSS has the potential to revolutionize the diagnosis of bladder cancer by offering real-time, personalized recommendations to clinicians. As AI technology becomes more explainable, its incorporation into clinical decision-making is likely to grow.
4) Telemedicine and remote diagnostics
AI-powered diagnostic tools for bladder cancer could be expanded to include telemedicine applications, enabling remote analysis of cystoscopy images or urine samples. This could enhance access to specialist care, especially in underserved regions. In recent years, clinical trials have started to integrate AI into the diagnosis of bladder cancer, focusing on how it can be incorporated into existing diagnostic workflows. These trials are designed todetermine whether AI systems, such as those analyzing cystoscopy images, can decrease the workload for healthcare providers and improve diagnostic accuracy. To tackle the variability among patient populations, several multicenter and international trials are currently in progress. These trials aim to validate AI models using larger and more diverse datasets, thereby improving their generalizability and robustness across different healthcare systems.
7. Clinical Trials and Validation of AI Models
The clinical validation of AI models for bladder cancer diagnosis is a critical step in transitioning from research to clinical practice. Approval and the adoption of AI tools in clinical settings necessitate rigorous trials and validation to confirm that these models are accurate, safe, and reliable when used with real-world patient data.
1) Prospective vs. retrospective studies
Most AI models, particularly in their developmental phases, are validated using retrospective datasets derived from historical patient data. This approach is useful for establishing an initial proof of concept and performance metrics, including sensitivity, specificity, and AUC. However, these models require prospective validation with new patient data in real-world clinical settings. Prospective clinical trials are crucial for assessing the performance of AI models when integrated into the intended clinical workflow and for addressing potential biases associated with relying solely on retrospective datasets.
Key clinical trials and studies are summarized in Table 7.
Key clinical trials and studies
2) Regulatory and ethical challenges in validation: data standardization
One of the significant challenges in validating AI models across multiple clinical centers is the lack of standardized data. Variations in imaging equipment, settings, and protocols across institutions can impact the performance of AI models. In the context of bladder cancer diagnostics, there can be substantial differences in cystoscopy images and urine cytology samples between hospitals. This requires extensive preprocessing and data harmonization before AI models can be effectively deployed on a large scale. Therefore, multicenter trials that utilize standardized datasets or datasets that have been harmonized through normalization techniques are essential to demonstrate the generalizability of AI models.
3) Real-world data and evidence
AI models must be tested with real-world data to confirm their effectiveness outside of controlled research settings. When AI tools are applied in clinical environments, they frequently encounter unexpected challenges, such as variations in patient demographics or disease presentations that were not accounted for in the training datasets. Regulatory bodies, including the FDA in the United States, have established guidelines for assessing AI tools using real-world evidence collected from clinical practice. This is especially crucial in the field of bladder cancer diagnostics, where AI models need to show consistent performance in everyday workflows.
4) FDA-approved AI models
As of now, only a few AI models for cancer diagnosis, including those for bladder cancer, have received FDA approval. These models are subject to rigorous testing, which includes preclinical validation and subsequent prospective clinical trials. For instance, the IDx-DR system, an AI tool designed for diabetic retinopathy, was among the first AI systems to gain FDA approval. While not specifically related to bladder cancer, the validation process of IDx-DR provides a blueprint for the validation of future AI systems aimed at diagnosing bladder cancer. IDx-DR underwent validation through a large, multicenter, prospective clinical trial that included thousands of patients. In this trial, the system’s performance was benchmarked against that of expert ophthalmologists, showing comparable accuracy. Similar trials will be essential for AI tools targeting bladder cancer to confirm their safety and efficacy in various clinical environments and among different patient demographics.
8. Reference Validated Studies
We developed a CNN algorithm and validated it using 2,102 cystoscopy images, achieving a high sensitivity of 89.7% and a specificity of 94%. Although this study was retrospective, it established a foundation for future prospective trials [11].
Previous research evaluated different CNNs (LeNet, AlexNet, and GoogLeNet) in bladder cancer diagnosis, achieving a high accuracy of 96.9%. This model has moved toward prospective validation in clinical settings [12].
Focused on the preoperative staging of bladder cancer, the study utilized MRI radiomics and SVM-based models, achieving an AUC of 0.9857. Their model exceeded human diagnostic accuracy in distinguishing non-muscle-invasive from muscle-invasive bladder cancer, with prospective trials currently in the planning stages [21].
9. Challenges in Clinical Validation
1) Lack of large-scale prospective trials
While many AI models have demonstrated exceptional performance in retrospective studies, very few have been subjected to the large-scale prospective trials required for clinical validation. Although retrospective validation is valuable, it does not fully account for the real-world variability in clinical workflows.
2) Ethical concerns
AI models must ensure that patient data is anonymized and that data usage adheres to ethical standards, particularly in terms of consent and privacy. There is a need for continuous ethical oversight, especially in multicenter trials that involve sensitive medical data.
3) Regulatory pathways
AI-based diagnostics, particularly for bladder cancer, must adhere to rigorous regulatory frameworks, including the FDA’s Software as a Medical Device approval process. Both prospective trials and real-world evidence are essential for these models to secure regulatory approval.
AI has demonstrated considerable potential in histopathological analysis, where ML algorithms support pathologists in examining tissue samples. Digital pathology, enhanced by AI, enables the automated identification and grading of bladder cancer in histopathological slides. Research has shown that AI algorithms can attain diagnostic accuracy on par with expert pathologists while also reducing interobserver variability.
The integration of AI with CDSS represents a significant advancement. AI enhanced CDSS can offer real-time recommendations to clinicians by analyzing patient data and medical literature. For example, AI algorithms are capable of suggesting diagnostic tests, interpreting results, and recommending treatment plans customized for individual patients. These systems have the potential to improve clinical decision-making, enhance diagnostic accuracy, and optimize patient management.
AI has also been applied to the monitoring and surveillance of bladder cancer patients. Automated systems are capable of analyzing follow-up imaging and urine samples to detect any recurrence or progression of the disease. AI-powered tools enable continuous monitoring, which facilitates timely interventions and enhances long-term patient outcomes.
LIMITATIONS
One of the primary challenges in using AI for bladder cancer diagnosis is the quality and standardization of data. High-quality, annotated datasets are crucial for training accurate AI models. Variability in imaging protocols, sample preparation, and annotation standards can impact the performance of AI algorithms. Therefore, efforts to standardize data collection and annotation practices are essential for the successful implementation of AI in clinical practice [10,11].
AI algorithms, particularly DL models, often function as “black boxes,” which obscures their decision-making processes. To gain the trust of clinicians and patients, it is crucial to ensure the interpretability and explainability of these AI models. Techniques like XAI are being developed to shed light on the reasoning behind AI algorithms’ conclusions, thereby enhancing transparency and accountability.
Integrating AI tools into existing clinical workflows presents practical challenges. It is essential to ensure seamless integration with EHRs and other clinical systems for the effective use of AI in routine practice. Critical factors for successful implementation include user-friendly interfaces, interoperability, and robust validation in real-world settings. The use of AI in healthcare introduces several ethical and legal challenges, such as concerns about patient privacy, data security, and algorithmic bias. It is crucial to ensure the ethical application of AI, safeguard patient data, and tackle biases within AI algorithms to maintain public trust and promote equitable healthcare.
The integration of AI and ML models into routine clinical practice is hindered by regulatory challenges. Numerous AI/ML-based medical devices are currently pending approval from the FDA and Conformite Europeenne. As the number of these technologies continues to grow, the delays in securing these essential approvals are also increasing [23,24].
Another potential limitation is the diverse attitudes and perspectives of patients toward AI which lead to varying levels of support, understanding, and acceptance of its use in routine clinical practice [25,26].
CONCLUSIONS
The use of AI in diagnosing bladder cancer shows significant potential for enhancing diagnostic precision, efficiency, and patient outcomes. Developments in image analysis, urine biomarker analysis, predictive analytics, and CDSS are revolutionizing the approach to bladder cancer diagnosis. Although there are challenges concerning data quality, interpretability, integration, and ethical issues, continuous research and development are facilitating the broader implementation of AI in clinical settings. The future of bladder cancer diagnosis will likely involve a synergistic integration of AI with conventional diagnostic techniques, promoting early detection, personalized treatment, and improved patient care.
Notes
Grant/Fund Support
This study received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.
Conflicts of Interest
The authors have nothing to disclose.
Author Contribution
Conceptualization: BA, CS, YMC; Data curation: JJ, YIL, JMP, SYY; Formal analysis: YIL; Funding acquisition: YMC; Methodology: YIL; Project administration: JMP, SYY; Visualization: JJ; Writing - original draft: BA, YMC; Writing - review & editing: BA, CS.