Artificial intelligence and predictive tools in non-muscle invasive bladder cancer: a narrative review of current insights and advances

Abstract

Background and Objective

Non-muscle invasive bladder cancer (NMIBC) is characterized by high recurrence rates and heterogeneous progression risk, making accurate diagnosis, risk stratification, and personalized management challenging. Conventional clinical scoring systems provide general guidance but often fail to fully capture tumor complexity and interpatient variability. This review summarizes current applications of artificial intelligence (AI) in NMIBC, focusing on diagnosis, prognostic, and clinical decision-making.

Methods

A comprehensive literature search was conducted in PubMed, Google Scholar, Embase and Scopus. Keywords related to AI and NMIBC including machine learning, deep learning, imaging, cystoscopy, radiomics, and computational pathology were used. Studies were independently screened, followed by full-text assessment for eligibility. A total of 35 English-language studies, published between January 2019 and March 2026, were included in the final qualitative synthesis.

Key Content and Findings

AI applications in NMIBC span cystoscopy, imaging, histopathology, and prognostic modeling, demonstrating high diagnostic and predictive performance. In cystoscopy, deep learning models achieve sensitivities ranging from 88% to 97% and specificities from 92% to 99%, with area under the curves (AUCs) up to 0.98–0.99. Real-time segmentation reports Dice coefficients between 74% and 93%, with processing times approximately 6–7 ms per image. In imaging, AI-based radiomics and deep learning applied to magnetic resonance imaging (MRI) and computed tomography (CT) provide AUCs ranging from 0.82 to 0.99, often outperforming conventional models. Multiparametric MRI achieves AUCs of 0.88–0.91 for recurrence prediction, while CT-based models reach up to 0.997 for differentiating NMIBC from muscle-invasive disease. Prognostic models using machine learning, including random survival forests and neural networks, demonstrate improved discrimination compared to traditional scores, with concordance indices up to 0.79–0.88, enabling more granular risk stratification. In histopathology, AI-driven analysis of whole-slide images achieves accuracies of 74–90% for recurrence prediction and AUCs up to 0.86, while identifying patients at significantly higher risk of progression or treatment failure.

Conclusions

AI enhances NMIBC management by enabling more precise, reproducible, and individualized diagnosis and risk assessment. The integration of multimodal data may improve clinical decision-making and support personalized treatment strategies, although further validation and standardization are required before widespread clinical implementation.

Introduction

Non-muscle invasive bladder cancer (NMIBC) represents a heterogeneous group of urothelial malignancies characterized by high recurrence rates and variable progression risk (1,2). Over the past two decades, multiple traditional clinical risk stratification tools have been proposed, including the European Organisation for Research and Treatment of Cancer (EORTC) (3), Spanish Urological Club for Oncological Treatment (CUETO) (4), European Association of Urology (EAU) (5), American Urological Association/Society of Urologic Oncology (AUA/SUO) (6), National Comprehensive Cancer Network (NCCN) 2024 (7), and International Bladder Cancer Group (IBCG) (8) scoring systems, providing general guidance for initial management and surveillance. However, these models often fail to capture the biological complexity of NMIBC and the heterogeneity of individual patient profiles, thereby limiting their predictive accuracy (3-8).

Recent advances in artificial intelligence (AI) offer promising opportunities to address the limitations of traditional diagnostic and risk assessment tools in NMIBC (9). By leveraging large, complex datasets from cystoscopy, imaging, digitized histopathology, and clinical variables, AI algorithms can detect subtle patterns imperceptible to human observers, improve diagnostic accuracy, refine risk stratification, and guide individualized therapeutic decisions (10-13). Specifically, machine learning and deep learning approaches, such as convolutional neural networks (CNNs), random survival forests, and transformer-based models, provide a robust framework for integrating these multimodal data sources, enabling more comprehensive and individualized clinical decision-making (14-16). Despite these promising developments, the rapid expansion of AI applications in NMIBC requires systematic evaluation to determine their clinical utility, reproducibility, and integration into routine practice (17,18).

The present review aims to provide a comprehensive synthesis of current evidence on AI in NMIBC, encompassing cystoscopy detection, advanced imaging analysis, histopathological assessment, and prognostic modeling. By highlighting both technical performance and potential clinical impact, this work seeks to inform the development and adoption of AI-driven strategies for precision management of NMIBC. We present this article in accordance with the Narrative Review reporting checklist (available at https://tau.amegroups.com/article/view/10.21037/tau-2026-0342/rc).

Methods

A comprehensive literature search limited to English-language publications was conducted in March 2026 to identify relevant studies on the use of AI in the management of NMIBC.

Search strategy

Relevant studies were identified through systematic searches of four electronic databases: PubMed, Google Scholar, Embase and Scopus, covering the period from January 2019 to March 2026. In PubMed, both Medical Subject Headings (MeSH) terms and free-text keywords were used in combination with Boolean operators (AND/OR) to ensure a comprehensive and sensitive search strategy. The following keywords were used in various combinations: “artificial intelligence”, “machine learning”, “deep learning”, “non-muscle invasive bladder cancer”, “bladder cancer”, “risk stratification”, “predictive modeling”, “precision medicine”, “computational pathology”, “digital pathology”, “medical imaging”, “radiomics”, “MRI”, “CT”, “cystoscopy”, and “virtual cystoscopy”. Additional relevant articles were identified through manual screening of reference lists of selected studies to ensure comprehensive coverage of the topic.

Study selection: inclusion and exclusion criteria

Studies were included if they met all of the following criteria: (I) original research articles; (II) evaluation of AI, machine learning, or deep learning methods applied to NMIBC; (III) focus on at least one clinically relevant domain, including diagnosis, detection, segmentation, risk stratification, prognostic modeling, imaging analysis, cystoscopy, or histopathology; and (IV) reporting of quantitative performance metrics [area under the curve (AUC), sensitivity, specificity, accuracy, Dice score, or survival-related outcomes]. Studies were excluded if they met any of the following criteria: (I) non-original publications (including reviews, editorials, expert opinions, guidelines, book chapters, and conference abstracts without full peer-reviewed manuscripts); (II) case reports or very small series without AI-based quantitative modeling; (III) studies not specifically focused on NMIBC or where NMIBC-specific results could not be extracted separately; and (IV) studies lacking sufficient methodological or performance data for meaningful synthesis. A summary of the literature search strategy and study selection process is presented in Table 1.

Table 1. The search strategy summary.

Items	Specification
Date of search	March 2026
Databases searched	PubMed, Google Scholar, Embase, Scopus
Search terms used	“Artificial intelligence”, “machine learning”, “deep learning”, “non-muscle invasive bladder cancer”, “bladder cancer”, “risk stratification”, “predictive modeling”, “precision medicine”, “computational pathology”, “digital pathology”, “medical imaging”, “radiomics”, “MRI”, “CT”, “cystoscopy”, “virtual cystoscopy”
Timeframe	January 2019 to March 2026
Inclusion and exclusion criteria	Inclusion: (I) original research articles; (II) evaluation of AI, machine learning, or deep learning methods applied to NMIBC; (III) focus on at least one clinically relevant domain, including diagnosis, detection, segmentation, risk stratification, prognostic modeling, imaging analysis, cystoscopy, or histopathology; (IV) reporting of quantitative performance metrics (AUC, sensitivity, specificity, accuracy, Dice score, or survival-related outcomes)
	Exclusion: (I) non-original publications (including reviews, editorials, expert opinions, guidelines, book chapters, and conference abstracts without full peer-reviewed manuscripts); (II) case reports or very small series without AI-based quantitative modeling; (III) studies not specifically focused on NMIBC or where NMIBC-specific results could not be extracted separately; and (IV) studies lacking sufficient methodological or performance data for meaningful synthesis
Selection process	Titles and abstracts were independently screened by two reviewers (P.E.G. and J.C.I.), and discrepancies were resolved through consensus discussion with two senior reviewers (J.P.S. and E.X.)

AI, artificial intelligence; AUC, area under the curve; NMIBC, non-muscle invasive bladder cancer.

Screening process

Study selection followed a structured two-stage screening process. First, titles and abstracts identified through the search were independently screened by two reviewers (P.E.G. and J.C.I.). Second, full-text articles meeting inclusion criteria or deemed potentially relevant were assessed independently by the same reviewers. Discrepancies were resolved through consensus discussion with two senior reviewers (J.P.S. and E.X.). A total of 35 studies were included in the final qualitative synthesis.

Data extraction and synthesis

Data were extracted using a standardized extraction framework, including: study design, dataset origin and characteristics, sample size, AI methodology (e.g., CNN, random forest, survival models, transformers), feature type (imaging, histopathology, clinical, or multimodal), training strategy, validation approach (internal and/or external validation), and reported performance metrics. Given the narrative nature of the review and the heterogeneity of included studies in terms of design, endpoints, and AI methodologies, no quantitative synthesis or meta-analysis was performed. Instead, findings were organized thematically according to clinical application domains (cystoscopy, imaging, prognostic modeling, and histopathology).

Critical appraisal

Although no formal risk-of-bias assessment tool (QUADAS-2 or PROBAST-AI) was systematically applied, all included studies were critically appraised with particular attention to methodological limitations, including retrospective design, dataset size and representativeness, class imbalance, risk of overfitting, absence of external validation, and generalizability to real-world clinical practice.

Results

AI applications for bladder cancer detection and segmentation in cystoscopy

Since 2019, AI has demonstrated remarkable potential for improving the detection, grading, and segmentation of BC from cystoscopy images (Table 2). Early approaches utilized CNNs to analyze real-time cystoscopy videos, achieving per-frame sensitivity of 88.2–90.9% and specificity of 98.6–99.0% for papillary tumors and flat lesions (19). In 2020, Ikeda et al. reported that a similar transfer learning-based model trained on 2,102 white-light images achieved an AUC of 0.98, with a sensitivity of 89.7% and a specificity of 94.0%, demonstrating stable performance across Ta–T2 stages as well as for small (<10% of image area) or flat lesions (20). Blue-light cystoscopy has also benefited substantially from AI, with models outperforming physicians by 15–40% for malignancy detection, tumor stage (T1/T2), and grade (sensitivities 88–96%, specificities 87–96%) (21). Multicenter AI systems have further demonstrated robust performance at large scale. An analysis of 69,204 images from 10,729 patients reported internal accuracy of 97.7%, sensitivity and specificity above 97.5%, and external validation accuracies ranging from 97.8% to 99.1%. Notably, processing efficiency was markedly improved, with 260 images analyzed in 12 seconds compared to 35–45 minutes for urologists (22).

Table 2. AI for bladder cancer detection, characterization, and segmentation in cystoscopy.

Studies	Design/data characteristics	AI model/type	Data used	Main results
Shkolyar et al., 2019 (19)	Retrospective single-center development cohort with separate prospective validation cohort; real-world cystoscopy and TURBT video data	CNN (CystoNet)	141 videos: 100 patients (development), 54 patients (limited external validation); ~2,700 tumor frames, 33,000 normal frames	Per-frame sensitivity 88–91%, specificity 98–99%; per-tumor sensitivity 95.5%; detected papillary tumors and CIS; real-time video overlay support
Ikeda et al., 2020 (20)	Rétrospective, single-center study; real-world white-light cystoscopy images. No external validation	CNN (GoogLeNet + transfer learning)	2,102 white-light images, 109 patients; tumor types: elevated 61%, flat 18%, mixed 21%; stages Ta–T2 including CIS	AUC 0.98; sensitivity 89.7%, specificity 94%; robust across stages and morphologies; small or flat lesions sometimes missed
Ali et al., 2021 (21)	Retrospective multi-center study; real-world bladder imaging dataset; external validation implied through multi-center data split (no independent external cohort)	CNN (MobileNetV2, ResNet50, VGG16, InceptionV3)	216 lesions from multi-center bladder images; tumor grades: low/high	Malignant detection: 91.8% sensitivity, 77–97% specificity; T-stage detection: mean 88% sensitivity, 96.6% specificity; CNNs outperformed physicians by 25–40%
Wu et al., 2022 (22)	Retrospective multicenter study with internal validation and multiple independent external validation cohorts; real-world cystoscopy images including routine clinical and TURBT cases.	CNN (CAIDS; PSPNet + ResNet101 + pyramid pooling)	69,204 images from 10,729 patients across 6 hospitals; multiple external datasets (up to 4 independent centers) with pathology-confirmed positives and biopsy/follow-up-confirmed negatives	Accuracy 97.7–99%, sensitivity 87.5–98.7%, specificity 97.5–99.6%; F1 0.939; diagnoses in 12 s vs. 35–45 min for humans; robust across tumor stages
Ikeda et al., 2021 (23)	Retrospective single-center study using real-world cystoscopy white-light images; no external validation	CNN (GoogLeNet + stepwise transfer learning)	2,102 images, 109 patients; 8:2 train/test	Sensitivity 95.4%, specificity 97.6%, AUC 0.98; higher accuracy for tumors >10% of image; faster than human observers
Yoo et al., 2022 (24)	Retrospective single-center study using real-world cystoscopy WLI and NBI; no external validation	AI-assisted SVM on color features (WLI/NBI)	900 images (300 tumor, 600 normal)	Sensitivity 95%, specificity 93.7%, accuracy 94.1%, AUC 0.974; good grading and location detection; WLI better than NBI
Varnyú et al., 2022 (25)	Retrospective single-center study; real-world WLI cystoscopiy images; pixel-wise semantic segmentation; no external validation	CNN (LinkNet/Encoder–Decoder)	2,578 images, 7 classes segmentation	Mean Dice 89.7%, malignant F-score 87%; real-time segmentation; distinguishes malignant, benign, and healthy tissue
Mutaguchi et al., 2022 (26)	Retrospective single-center study; real-world TURBT video-derived images; no external validation	Dilated U-Net	1,790 TURBT images, 120 patients; classes: Ta, T1, T2, CIS	Pixel-wise sensitivity 84.9%, specificity 88.5%, PPV 86.7%, Dice 83%; improved over standard U-Net; segments tumors accurately
Zhang et al., 2023 (27)	Retrospective single-center study; real-world WLI/TURBT cystoscopy images; no external validation	ACS (attention-based segmentation)	1,000 images, 237 patients	Dice 82.7%, Mean IoU 69%; attention modules improved accuracy; multi-scale tumor segmentation with fewer errors
Jia et al., 2023 (28)	Retrospective single-center study; real-world WLI cystoscopy images; no external validation	CystoNet-T (ResNet50 + transformer)	611 images, 67 patients	F1 96.4%, AP 91.4%, recall 97.3%, precision 95.6%; better than Faster R-CNN and YOLO; accurate multi-scale tumor localization
Kim et al., 2024 (29)	Retrospective, multicenter study; real-world cystoscopy images; no external validation	AI-assisted software (INF-M01)	1,890 images; 486 cancer, 1,404 normal; 5 validation sets	Sensitivity 95–97%, specificity 88–96%, Dice 87–92%; comparable to experienced urologists; identifies suspicious regions autonomously
Ye et al., 2025 (30)	Prospective-retrospective single-center study; real-world cystoscopy/TURBT videos; no external validation	Semantic segmentation (HRNetV2 + FCN)	102 videos, 94 patients; 33,657 annotated frames	Test sensitivity 91.6%, precision 91.3%, mean Dice 80%; post-processing improved detection; false negatives mainly small/atypical lesions

ACS, attention-based cystoscopic image segmentation; AI, artificial intelligence; AP, average precision; AUC, area under the curve; CIS, carcinoma in situ; CNN, convolutional neural network; Dice, Dice similarity coefficient; F1, f1 score (harmonic mean of precision and recall); FCN, fully convolutional network; IoU, intersection over union; NBI, narrow-band imaging; PPV, positive predictive value; TURBT, transurethral resection of bladder tumor; WLI, white-light imaging.

More advanced strategies have subsequently been developed to further enhance diagnostic performance. For instance, stepwise transfer learning strategies increased sensitivity, particularly for small or flat lesions, with 95.4% sensitivity, 97.6% specificity, and AUC 0.98 (23). Furthermore, an AI-assisted diagnostic device applied to 900 cystoscopy images achieved 95.0% sensitivity, 93.7% specificity and 94.1% overall accuracy. The system also analyzed tumor color using average red, green and blue values to differentiate benign, low-grade, and high-grade tumors, with a Dice coefficient of 74.7% for lesion localization. In addition, white-light images achieved over 98% accuracy for benign versus malignant differentiation and over 90% for differentiating chronic non-specific inflammation from carcinoma in situ, highlighting the potential for real-time AI-assisted cystoscopy and improved tumor characterization (24).

In parallel, real-time semantic segmentation has been explored using pixel-wise models. Dice scores ranged from 74.7% to 93%, with high pixel-wise F-scores and precision (83–93%) (25,26). Some models achieved the highest overall accuracy (Dice 92.9%, F-score ≈91%), while others were optimized for real-time processing, analyzing each image in approximately 6.7 ms (25). The use of dilated convolution layers and advanced network architectures further improved segmentation performance, particularly for elevated and mixed lesions, demonstrating that these AI approaches can accurately delineate tumors of varying morphologies and effectively support clinical decision-making (26). Moreover, incorporating attention mechanisms improved accuracy, reduced false positives and negatives, and refined tumor boundaries, reaching 82.7% Dice and 69% mean intersection over union (MioU) (27). A model combining CNNs with global attention mechanisms achieved 97.3% recall, 95.6% precision, F1-score 96.4%, thereby enhancing multi-scale tumor detection and localization (28). Clinical studies have further confirmed the reproducibility and reliability of AI analysis. For example, on 1,890 images, AI achieved 97.3% sensitivity, 92.1% specificity, 93.4% accuracy, and a mean Dice coefficient of 0.903, comparable or superior to four experienced urologists (29). Similarly, another AI model applied to cystoscopy videos demonstrated overall sensitivity of 91.6% and precision of 91.3%, with maximum performance on high-resolution frames (sensitivity 94.8%, precision 94.4%, Dice 84.7%) and acceptable performance even on low-resolution frames (sensitivity 75.6%, precision 74.8%, Dice 56.6%) (30).

AI applications in imaging for diagnosis and risk stratification of NMIBC

AI has also increasingly demonstrated its potential to enhance non-invasive diagnosis, risk stratification, and prognostication in NMIBC using both magnetic resonance imaging (MRI) and computed tomography (CT) imaging (Table 3). In the MRI domain, multiparametric MRI (mp-MRI) combined with radiomics and deep learning approaches has been extensively investigated. A retrospective study of 191 NMIBC patients showed that a hybrid model integrating clinical, radiomics, and deep learning data from both intratumoral and peritumoral regions achieved high predictive performance for 5-year recurrence, with AUCs of 0.88 and 0.91 in validation and testing cohorts, respectively, significantly outperforming conventional EORTC risk models (31). Similarly, in a cohort of 183 patients, radiomic features extracted from T2-weighted, apparent diffusion coefficient (ADC), and dynamic contrast-enhanced (DCE) sequences, combined with clinical variables, were used to train machine learning models; the support vector machine (SVM) achieved AUCs of 0.89 in validation, with external datasets confirming generalizability (32).

Table 3. AI applications in imaging (MRI and CT) for diagnosis and risk stratification of NMIBC.

Studies	Design/data characteristics	AI model/type	Data used	Main results
Huang et al., 2025 (31)	Retrospective single-center study; real-world mp-MRI radiomics dataset; no external validation	Clinical model (AutoML), Radiomics (Rad), Deep learning (ResMANet), Hybrid (CRDL)	191 NMIBC patients; mp-MRI images; 113 radiomics features + clinical scores	CRDL model best: AUC 0.91 (test); better than EORTC; improved recurrence prediction
Chen et al., 2025 (32)	Retrospective single-center study; real-world mp-MRI radiomics dataset; internal + external validation	Clinical-radiomics + SVM	183 NMIBC patients (57 recurrence, 126 non-recurrence); mpMRI (T2WI, ADC, DCE); external validation 54 patients/hospital	SVM model: AUC 0.89, accuracy 88.9%, sensitivity 67%, specificity 94%; nomogram with 4 radiomics + 5 clinical features; robust 2-year recurrence prediction
Huang et al., 2026 (33)	Retrospective multicenter study; real-world clinical imaging data; cross-center validation; no external independent cohort	DADCNet (CNN + domain adaptation + contrastive learning)	Multi-center MRI dataset, NMIBC + MIBC; 10-fold cross-validation; `cross-center (leave-one-center-out) evaluation	Accuracy 95%, F1 0.955, AUC 0.991; outperformed ResNet, EfficientNet, Transformers; interpretable predictions focused on tumor and muscle layers
Song et al., 2025 (34)	Retrospective single-center study; real-world clinical dataset; internal split (train/validation/test); no external validation	TM3DConvNet (Transformer + 3D CNN)	184 patients (136 NMIBC, 48 MIBC); multi-phase DCE-MRI	AUC 0.82–0.89; sensitivity 0.90; outperformed VGG16, ResNet101, DenseNet121, ViT, and VI-RADS; precise spatiotemporal tumor focus
Zou et al., 2025 (35)	Retrospective single-center study; real-world MRI dataset; 5-fold cross-validation; no external validation	MVSD (multi-view self-distillation, 3D ResNet50)	615 patients (443 NMIBC, 172 MIBC); multi-view MRI slices; segmented bladder & tumor	AUC 0.927, accuracy 0.88, sensitivity 0.90, specificity 0.87; multi-view fusion + self-distillation improved classification
Yu et al., 2024 (36)	Retrospective multicenter study; real-world MRI dataset; internal and external cohort validation	Multisequence fusion DL (pNMI-RNet + CPA-Unet + DMF)	436 patients (288 NMIBC, 148 MIBC); T2WI, DWI, DCE	AUC 0.913–0.928; accuracy 0.831–0.869; sensitivity 0.69–0.75; outperformed VGG16, ResNet50, MobileNet; comparable to radiologist; improved MIBC prediction
Ye et al., 2023 (37)	Retrospective multicenter study; real-world clinical MRI cohort; internal and independent external validation	Semi-automatic segmentation (U-Net) + Radiomics (SVM)	119 patients, 160 tumors; T2 MRI; semi-automatic VOIs	Dice 0.80–0.84; Radiomics AUC 0.892–1.0; semi-automatic faster than manual (35 vs. 92s per lesion); equivalent accuracy
Yang et al., 2021 (38)	Retrospective single-center study; real-world CT dataset; no external validation	Small DL-CNN + 8 pretrained CNNs	369 patients (NMIBC 249, MIBC 120); 1200 contrast-enhanced CT images	Best: VGG16 AUROC 0.997, accuracy 0.939, sensitivity 0.889, specificity 0.989; other models varied
Jin et al., 2025 (39)	Retrospective single-center study; real-world CT dataset; no external validation	DLCS model (DenseNet121 + Rad + Clinical), Radiomics model, DL model	181 NMIBC patients; enhanced CT; 2.5D slices; 2048 DL + 837 radiomics features	DLCS + RF classifier best: AUC 0.894; Rad & DL scores predicted recurrence-free survival (HR Rad 38.90, HR DL 1.17); good calibration and clinical utility

ADC, apparent diffusion coefficient; AI, artificial intelligence; AUROC/AUC, area under the receiver operating characteristic curve; CT, computed tomography; DCE, dynamic contrast-enhanced imaging; Dice, Dice similarity coefficient; DL, deep learning; DWI, diffusion-weighted imaging; F1, F1 score; HR, hazard ratio; MIBC, muscle-invasive bladder cancer; MP-MRI, multiparametric MRI; MRI, magnetic resonance imaging; NMIBC, non-muscle-invasive bladder cancer; Rad, radiomics; SVM, support vector machine; T2WI, T2-weighted imaging; VOI, volume of interest.

Subsequently, deep learning approaches further improved MRI-based NMIBC detection and classification. A domain-adaptive deep contrastive network (DADCNet) achieved precision and F1-scores of 95.5% and an AUC of 0.99 for NMIBC classification, with cross-center validation confirming its robustness and clinical applicability (33). Multi-phase DCE-MRI data also enabled accurate prediction of NMIBC versus MIBC, with sensitivity up to 0.90 and AUC of 0.82, outperforming conventional imaging assessment (34). Multi-view fusion models based on three-dimensional (3D) T2-weighted sequences demonstrated similar robust results, achieving AUCs up to 0.927, accuracy of 0.88, and sensitivity above 0.90 in cohorts exceeding 600 patients (35). Another mp-MRI study confirmed strong internal and external validation performance, with AUCs exceeding 0.91 and accuracy above 0.83, highlighting the reproducibility and reliability of MRI-based AI models for NMIBC detection (36). Beyond detection, semi-automatic MRI lesion segmentation has been explored to streamline radiomics workflows. These approaches achieved high accuracy (AUC 0.89–1.00) while significantly reducing segmentation time compared to manual approaches, reinforcing AI’s efficiency in clinical pipelines (37).

CT imaging has also benefited from AI applications in NMIBC. Using contrast-enhanced CT, deep learning CNNs differentiated NMIBC from MIBC with AUC values up to 0.997 and accuracy of 0.939, demonstrating the potential of CT as a non-invasive diagnostic tool, particularly when leveraging pretrained architectures such as VGG16 (38). Beyond structural classification, CT-based radiomics combined with deep learning enabled molecular and prognostic assessment in 181 NMIBC patients, predicting HER2 expression and recurrence risk. A combined AI model integrating radiomics, deep learning, and clinical data achieved an AUC of 0.894 and provided independent prognostic information for recurrence-free survival (RFS), highlighting the dual diagnostic and prognostic value of CT-based AI (39).

AI applications for prognostic modeling and personalized risk stratification in NMIBC

Recent advances in AI have significantly improved risk stratification in NMIBC, addressing the limitations of traditional models previously described (Table 4). A first step in this evolution was the development of deep learning-based prognostic tools integrating established clinicopathological variables into more flexible modeling frameworks. In a multicenter study, a deep learning survival model trained on over 3,500 patients enabled individualized prediction of RFS and progression-free survival (PFS) by incorporating variables such as age, tumor stage and grade, size, multiplicity, and intravesical treatments. This approach improved predictive performance compared with conventional scores, with concordance indices of 0.88 for PFS and 0.65 for RFS in external validation, particularly enhancing progression risk estimation (40). Building on these approaches, machine learning models have further enhanced the understanding of key prognostic determinants and their impact on survival outcomes. In a large cohort of 1,510 patients with high-risk NMIBC treated with BCG, a random survival forest analysis identified time to progression and age as the strongest predictors of overall survival (OS), outperforming traditional regression models (C-index 0.81 in training and 0.73 in validation). Importantly, this study demonstrated a strong correlation between progression and mortality and highlighted that the timing of progression is critical: patients with early progression had significantly worse oncological outcomes. A survival tree-based model stratified patients into five risk groups with highly distinct survival probabilities, with 7-year OS ranging from over 98% in the lowest-risk group to less than 30% in the highest-risk group, emphasizing the prognostic value of disease dynamics beyond static baseline features (41). More recently, large-scale AI-driven models such as PROGRxN-BCa have further advanced prognostic accuracy by leveraging contemporary, international datasets including 12,659 NMIBC patients. Using a random survival forest algorithm and incorporating 14 routinely available clinicopathological variables at diagnosis, this model demonstrated superior discrimination compared with the EAU risk calculator (C-index 0.79 vs. 0.71, P<0.001). Importantly, it enabled more granular substratification within intermediate- and high-risk groups, identifying subpopulations with markedly different 5-year progression risks, for example, ranging from approximately 2% to 15% within intermediate-risk patients, thereby improving patient selection for treatment intensification or closer surveillance (42). Large population-based studies have also highlighted the utility of AI in predicting long-term outcomes in T1 NMIBC. Using data from 32,060 patients treated with transurethral resection of bladder tumor (TURBT), artificial neural networks (ANNs) were developed to predict 5-year cancer-specific survival (CSS) and OS based on demographic, clinical, and tumor characteristics. Significant predictors included age, tumor grade and size, histology variant, prior NMIBC, race, marital status, and income. In validation, the ANN achieved accuracies of 79.4% for CSS and 69.3% for OS, with AUCs of 72.5% and 73.4%. High-risk features included high-grade, large, or histology variant tumors, older age, and black race, while protective factors included marriage and higher income, demonstrating that ANN-based models can effectively support prognostic stratification and guide clinical decision-making in NMIBC (43). Finally, AI has been used to integrate both patient-related and tumor-specific factors into predictive models for pathological upstaging after TURBT. In a study of 380 NMIBC patients, supervised machine learning algorithms outperformed conventional logistic regression, achieving an AUC of 0.796. Tumor-specific factors [tumor (T) stage, high grade, lymphovascular invasion] and patient-related determinants (general condition, renal function, nutritional status) were identified as independent predictors of upstaging. Decision curve analyses demonstrated better probabilistic accuracy and higher net clinical benefit, highlighting AI’s ability to provide holistic, individualized prognostic assessment that goes beyond traditional models and static risk scores (44).

Table 4. AI-based prognostic models for risk stratification and outcome prediction in NMIBC.

Studies	Design/data characteristics	AI model/type	Data used	Main results
Jobczyk et al., 2022 (40)	Multicentrique retrospective study; real-world clinical cohort; external validation included	DeepSurv (deep neural network Cox proportional hazards model), compared to EORTC & CUETO clinical models	Training 3,570 patients (2,557 after exclusion), external validation 322 patients; clinical variables (age, sex, T stage, grade, tumor number/size, BCG/MMC treatment), follow-up up to 10 years	DeepSurv model improved prediction over EORTC/CUETO; validation C-index: 0.6508 (RFS), 0.8814 (PFS); extended model up to 0.8764 (PFS) with MMC; better long-term prognostic accuracy and no overfitting; enables personalized recurrence and progression prediction
Porreca et al., 2024 (41)	Multicenter retrospective study; real-world High Risk NMIBC cohort (18 institutions); internal validation only, no external cohort	RSF model + Survival tree + Cox regression comparison	1,510 high-risk T1 NMIBC patients from 18 institutions, treated with TURBT + BCG; 80% training/20% test; clinical, biological and tumor variables (21 covariates), median follow-up 49 months	RSF showed good predictive performance: C-index 0.814 (training), 0.726 (test), IBS 0.060–0.092; time to progression and age most important predictors; survival tree identified 5 risk groups with distinct OS; ML model improved risk stratification and prognostic prediction
Kwong et al., 2026 (42)	Multicenter retrospective study; real-world NMIBC cohort; external validation (30 institutions)	RSF-based AI model (PROGRxN-BCa), compared to EAU risk calculator and LASSO Cox model	12,659 NMIBC patients (training 3,324 from 4 Canadian centers; external testing 9,335 from 30 international institutions); clinical and pathological variables (14 features), follow-up >5 years	PROGRxN-BCa outperformed EAU: C-index 0.79 vs. 0.71 (external test, P<0.001); good calibration and higher net benefit; robust across subgroups; improved risk stratification (5-year progression 2–31%); superior to traditional and prior AI models
Ślusarczyk et al., 2023 (43)	Population-based retrospective study; real-world data; no external validation independent cohort	ANN vs. LR and Cox proportional hazards models	SEER database (2004–2015); 32,060 T1 NMIBC patients treated with TURBT; training (70%) and validation (30%) cohorts; demographic, clinical, pathological, and survival data	ANN and LR showed similar performance for 5-year survival prediction; CSS prediction AUC ~72–73% and accuracy ~79%; OS prediction AUC ~73–74% and accuracy ~69%; key predictors: age, tumor grade, size, histology, and sociodemographic factors; low overfitting and large-scale population-level validation
Özden et al., 2026 (44)	Single-center retrospective study; real-world clinical cohort; no external validation	Multiple ML models (SVM, Random Forest, XGBoost, Naive Bayes, MLP, LASSO logistic regression) vs. conventional logistic regression	380 NMIBC patients [2010–2020]; comprehensive clinical, pathological, laboratory, inflammatory, nutritional, and frailty variables; 5-year follow-up	ML models outperformed traditional logistic regression (AUC 0.702); best performance with SVM (AUC 0.796) and LASSO (AUC 0.769); key predictors: tumor stage, grade, lymphovascular invasion, frailty, renal function; ML improved discrimination and clinical decision utility for predicting pathological upstaging

AI, artificial intelligence; ANN, artificial neural network; AUC, area under the curve; BCG, Bacillus Calmette-Guérin; C-index, concordance index; CSS, cancer-specific survival; CUETO, Spanish Urological Club for Oncological Treatment; EAU, European Association of Urology; EORTC, European Organisation for Research and Treatment of Cancer; IBS, integrated Brier score; LASSO, least absolute shrinkage and selection operator; LR, logistic regression; ML, machine learning; MLP, multilayer perceptron; MMC, mitomycin C; NMIBC, non-muscle-invasive bladder cancer; OS, overall survival; PFS, progression-free survival; RFS, recurrence-free survival; RSF, random survival forest; SEER, Surveillance, Epidemiology, and End Results; SVM, support vector machine; TURBT, transurethral resection of bladder tumor; XGBoost, extreme gradient boosting.

AI applications in histopathology for predicting and managing NMIBC

Multiple studies have demonstrated that AI-driven analysis of digitized histopathology can extract quantitative features from tumor cells and their microenvironment, providing objective, reproducible, and clinically actionable insights that outperform conventional risk models (Table 5).

Table 5. AI in digital histopathology for prognosis and therapeutic decision-making in NMIBC.

Studies	Design/data characteristics	AI model/type	Data used	Main results
Tokuyama et al., 2022 (45)	Single-center retrospective study; image-based real-world clinical cohort; no external validation	SVM and Random Forest on nuclear morphology	125 NMIBC patients; 877 ROI images, ~1,008,502 nuclei, 960 features; 2-year follow-up	SVM: 100% training accuracy, 83.8% ROI, 90% patient test accuracy; RF: 74.9% ROI, 86.7% patient; nuclear features accurately predict early recurrence
Lucas et al., 2022 (46)	Multicenter retrospective cohort; real-world data; no external validation	Deep learning: VGG16 + bidirectional GRU	359 patients 1-year/281 patients 5-yaer; H&E slides (~5.5M patches 1 year, ~4.4M patches 5 year)	1-year RFS: 65% (AUC 0.62), 5-year RFS: 74% (AUC 0.76); deep learning with clinical data outperformed clinical data alone; combining histopathology and clinical info improves recurrence prediction
Wang et al., 2024 (47)	Multicenter retrospective cohort; real-world data; no external validation independent cohort	Deep learning + RF: ResNet50, histogram + TF-IDF aggregation	210 Ta/T1 patients; 210 WSI, 114,462 patches; follow-up ≥1 yr	Test: AUC 0.86; sensitivity 90%, specificity 65%, accuracy 81%; Grad-CAM highlights predictive regions; aggregation improves recurrence prediction
Chang et al., 2025 (48)	Multicenter retrospective cohort; real-world HG Ta NMIBC; no external validation; comparison with clinical risk models (AUA/EAU)	CHAI: tumor morphology & microenvironment biomarkers	269 high-grade Ta NMIBC; median follow-up 32 months	Biomarkers associated with recurrence HR 2.23 and MIBC progression HR 4.55; outperformed AUA and EAU risk models; reclassified ~70% of patients
Packiam et al., 2026 (49)	Prospective multicenter real-world observational implementation study; no external validation	CHAI: Vesta Bladder BCGPredict and Risk Stratify	105 high-grade NMIBC; 40× slides; pre/post physician surveys	Clinical management changed in 67% of cases; intravesical therapy agent changed in 63 cases; biomarkers guided decisions in 68% of cases
Lotan et al., 2025 (50)	Multicenter retrospective real-world cohort study with independent external validation across geographically distinct institutions	CHAI deep learning: signatures for recurrence, progression, BCG resistance, cystectomy-free survival	944 high-risk NMIBC patients; 12 centers; 40× H&E slides	Stratified patients: recurrence HR 2.08–2.23, progression HR 3.24–3.87, BCG resistance HR 2.21–2.31; AUC 0.98–0.99; robust across scanners and magnifications
Jiang et al., 2025 (51)	Retrospective multicenter real-world cohort study; 5 hospitals; 4 external validation cohorts	ERPM & TRPM: ensemble deep learning, multi-instance (DSMIL)	1,275 NMIBC; 4,395 WSI H&E ± IHC (P53, CK20, Ki67); 1,135,413 patches	ERPM AUC 0.837; stratifies high/low recurrence risk; TRPM: accuracy 84.1%, sensitivity 69%, specificity 88.3%; combined ERPM + TRPM: three risk levels (low, high, very high)
Jansen et al., 2020 (52)	Multicenter retrospective real-world pathology cohort, routine clinical TURBT specimens [2000–2016]; no external validation cohort	U-Net for segmentation + VGG16 for grading	328 NMIBC samples; 1.2 M patches; 60% train, 20% validation and 20% test	U-Net detects urothelium; VGG16 κ 0.48, accuracy 74%; 76% low-grade, 71% high-grade correctly graded; comparable to pathologists
Parrao et al., 2026 (53)	Single-center retrospective real-world diagnostic cohort; no external validation	CNN on Masson trichrome biopsy images (NMIBC vs. MIBC)	32 patients; 702 images (22 NMIBC, 10 MIBC; Ta/T1/T2/T3)	Training accuracy 95.2%, validation 90.1%; AUC 0.8; sensitivity 90%, specificity 91%, PPV 82%, NPV 95%; detection of muscle invasion significant (P=0.02); robust after adjusting for clinical variables

AUA, American Urological Association; AUC, area under the curve; BCG, Bacillus Calmette-Guérin; CHAI, computational histology artificial intelligence; CNN, convolutional neural network; EAU, European Association of Urology; ERPM, early recurrence predictive model; GRAD-CAM, gradient-weighted class activation mapping; GRU, gated recurrent unit; H&E, hematoxylin and eosin; HG, high-grade; HR, hazard ratio; IHC, immunohistochemistry; MIBC, muscle-invasive bladder cancer; NMIBC, non-muscle-invasive bladder cancer; NPV, negative predictive value; PPV, positive predictive value; RF, random forest; RFS, recurrence-free survival; ROI, region of interest; SVM, support vector machine; TF-IDF, term frequency-inverse document frequency; TRPM, treatment response predictive model; TURBT, transurethral resection of bladder tumor; VGG16, 16-layer convolutional neural network (ImageNet pretrained); WSI, whole-slide image.

Several studies have highlighted the value of AI for predicting disease recurrence or progression. Tokuyama et al. reported the results of a study involving 125 NMIBC patients, in which hematoxylin-and-eosin-stained slides were digitized as whole-slide images and analyzed using AI-based algorithms. Nuclear segmentation was performed to extract 79 quantitative features, including nuclear shape and intranuclear texture, which were aggregated to capture heterogeneity and pleomorphism. Machine learning classifiers, including SVMs and random forests, predicted recurrence within two years, achieving 83.8% accuracy at the region level and 90% accuracy at the case level, highlighting the potential of nuclear morphometry combined with AI to predict early recurrence beyond traditional clinical risk stratification (45). Building on this approach, Lucas and colleagues analyzed 840 NMIBC patients to develop a deep learning-based model for recurrence prediction. Digitized histopathology slides were divided into small regions, from which features were extracted and combined with clinical variables in a neural network. The model predicted 5-year RFS with 74% accuracy (AUC 0.76), significantly outperforming conventional multivariable logistic regression models (AUC: 0.57–0.58). This study demonstrated that AI could integrate complex spatial tissue patterns with patient-level data, enhancing prognostic precision beyond standard clinical models (46). Similarly, a study of 210 NMIBC patients applied a deep learning framework to digitized tumor slides, generating patch-level predictions aggregated with histogram-based methods. A random forest classifier achieved 81% accuracy, 90% sensitivity, and 65% specificity (AUC 0.86), illustrating the value of AI-driven pathomics for risk stratification and clinical decision support (47). The recent Computational Histology Artificial Intelligence (CHAI) platform further exemplifies the clinical impact of AI in NMIBC. Applied to 269 BCG-naïve patients with high-grade Ta disease, CHAI quantified tumor cell morphology and microenvironment features from digitized slides to generate biomarkers predictive of disease recurrence and progression. CHAI effectively stratified patients, with hazard ratios (HRs) of 2.23 for high-grade RFS and 4.55 for PFS, outperforming traditional AUA and EAU risk models. Importantly, the AI component alone remained independently prognostic, enabling reclassification of patients more accurately than guideline-based schemes (48). A real-world observational study of 105 high-grade NMIBC cases demonstrated that CHAI influenced clinical management in 67% of patients, including changes in surgical versus intravesical therapy and selection of intravesical agents. Physicians reported using CHAI biomarker results in 68% of cases, highlighting the practical utility of AI for personalized treatment decisions (49).

Beyond recurrence prediction, AI has also been leveraged to forecast response to intravesical therapy. Thus, in a multicenter study of 944 high-risk NMIBC patients treated with TURBT and BCG, whole-slide images were analyzed to extract 600 microenvironmental features, including cellular pleomorphism, mitotic activity, immune infiltration, and tumor aggressiveness. AI models accurately predicted high-grade RFS, progression to MIBC, BCG-unresponsive disease, and cystectomy risk. The recurrence risk model classified 31.5% of patients as high-risk for high-grade recurrence, compared with only 14% identified by conventional EORTC classification. The progression risk model identified patients with a 3.87-fold higher risk of progression to muscle-invasive disease. These results emphasize that AI-derived histopathologic biomarkers provide independent prognostic value and support personalized management decisions (50). Further extending AI applications, a multicenter study of 1,275 NMIBC patients developed an ensemble deep learning model to predict early recurrence and a complementary model for BCG response. Over 4,000 digitized slides were processed to extract features at both patch and case levels, integrated with clinical data. The recurrence model achieved an AUC of 0.837, identifying high-risk patients with a six-fold higher risk of early recurrence (HR 6.53). The BCG response model predicted unresponsiveness with 84.1% accuracy, 69% sensitivity, and 88.3% specificity. Together, these models allowed stratification into low-, high-, and very high-risk groups, demonstrating the capacity of AI to enhance prognostic precision and guide personalized treatment strategies (51).

Finally, AI has also demonstrated utility in tumor grading and detection of muscle invasion. An automated grading system was developed using 328 NMIBC specimens from 232 patients across three centers. High-resolution digitized slides were annotated by uro-pathologists observers to identify tumorous and non-atypical urothelium, and AI models graded the tumors according to World Health Organization (WHO) criteria. The system achieved moderate agreement with consensus pathologists (κ=0.48) and accuracies of 76% for low-grade and 71% for high-grade tumors, successfully managing variability in staining and artifacts (52). In parallel, AI has improved the detection of muscle invasion. Using 702 Masson’s trichrome-stained biopsy images from 32 patients, a CNN distinguished NMIBC from MIBC by highlighting collagen and muscle fibers. The model achieved 95.2% accuracy in training and 90.1% in validation, with 90% sensitivity and 91% specificity. Multivariate analysis confirmed that the model’s probability scores independently predicted muscle invasion, outperforming clinical variables such as tumor grade, size, and weight, while reducing inter-observer variability in staging (53).

Discussion

In oncology, AI is increasingly providing the ability to analyze complex, high-volume data, reveal subtle patterns imperceptible to humans, and generate predictive guidance for patient treatment (54). In NMIBC, AI has moved beyond a supportive role to become a promising tool in the development of personalized care, combining cystoscopy, advanced imaging, digitized histopathology, and biological data to refine prognostic stratification (9,13).

In this comprehensive review, AI-assisted approaches in NMIBC demonstrate high and reproducible performance across cystoscopy, imaging, and histopathology, offering significant potential to improve diagnosis, risk stratification, and clinical decision-making. Thus, AI-assisted cystoscopy achieves sensitivities above 90% and segmentation performance comparable or superior to that of urologists, standardizing diagnosis, reducing inter-observer variability and accelerating analysis (19-30). Advanced imaging, including mp-MRI and CT scans, complements this analysis by distinguishing NMIBC from MIBC, characterizing tumor morphology, intratumoral heterogeneity, and peritumoral interactions, predicting recurrence at 3–5 years with AUCs between 0.78 and 0.88, and identifying molecular biomarkers such as HER2, providing a comprehensive framework for personalized care (31-39). Multivariable prognostic models that incorporate tumor characteristics, patient-specific variables, and dynamic disease features outperform traditional scoring systems in predicting RFS, PFS, and OS, and provide more accurate risk stratification to guide personalized surveillance, treatment, and overall patient management (40-44). Specifically, the recent PROGRxN-BCa tool, which incorporates 14 clinicopathological variables and is trained on thousands of patients, allows fine substratification of intermediate- and high-risk groups and enables dynamic risk re-evaluation according to BCG exposure and individual patient characteristics (42). Finally, AI-assisted analysis of digitized histopathology slides extracts fine-grained features of nuclear morphology and the tumor microenvironment, including immune infiltration, mitotic activity, and nuclear pleomorphism. These data improve recurrence prediction and therapeutic stratification, reclassifying patients into appropriate risk groups and identifying profiles associated with BCG response or progression risk (45-53). Taken together, these modalities provide a multidimensional approach capable of better capturing the complexity of NMIBC and improving personalized clinical decision-making.

Beyond diagnosis and prognosis, AI addresses practical challenges in NMIBC management. In cystoscopy, AI evaluates image quality in real time, ensures complete bladder coverage, and guides lesion detection, even in difficult areas such as the bladder neck or when urine is turbid (55). Advanced imaging complements this assessment by providing non-invasive, high-resolution tumor characterization, facilitating early detection and accurate evaluation of tumor extent. In the context of emerging bladder-sparing strategies, AI applied to cystoscopy and imaging could significantly help differentiate tumor lesions from inflammatory changes in patients with inflamed, heavily resected, or previously treated bladders, enabling targeted biopsies and reducing the risk of under- or overtreatment (24). Moreover, the recent evolution of NMIBC treatment, which now includes intensified systemic immunotherapy (56-59), makes precise patient stratification essential. Traditional clinical scores are no longer sufficient, and AI emerges as a strategic tool that integrates clinical, biological, radiomic, and histopathological data to improve risk prediction and optimize allocation of intensive therapies.

Clinical implementation of AI raises practical, ethical, and economic considerations. Initial costs remain high due to infrastructure and algorithm development (60,61); however, AI can streamline workflows, reduce manual review times, and optimize resource allocation (62,63). For instance, AI-assisted analysis of electronic medical records effectively identifies high-risk NMIBC patients and predicts BCG response using transformer-based natural language processing (NLP) models, achieving F1 scores up to 0.98 for time to recurrence and 0.94 for time to progression, significantly reducing manual workload (64). NLP models have also extracted key risk factors from unstructured clinical notes to identify high-risk patients accurately (65). Ethical considerations require that AI models remain transparent, interpretable, and validated in diverse populations, while ensuring data confidentiality, security, and informed consent (66-68). Environmental sustainability is also critical, as large-scale deep learning consumes substantial energy; algorithm optimization, efficient cloud use, and limiting redundant training remain essential strategies (69). Finally, interoperability with hospital systems, clinician training, and regulatory approval are crucial for safe and effective adoption (70-72).

To fully realize AI’s potential, developers must create multimodal models that integrate clinical, histopathological, radiomic, and molecular data, ideally trained on international cohorts to improve robustness and generalizability (73). Deep neural networks and SVMs dominate the field, capable of handling high-dimensional data, while hybrid approaches maintain interpretability while maximizing predictive power (74). Integration into clinical workflows through clinician-friendly interfaces and electronic health records, along with regulatory “sandbox” frameworks, allows progressive validation of these tools in a secure, compliant environment (75).

Thus, AI represents a promising emerging tool with the potential to significantly improve NMIBC management. By combining cystoscopy, advanced imaging, computational histopathology, and biological data, it provides a truly personalized and dynamic approach, recalculating risk at each visit and optimizing therapeutic strategies. The next step involves clinical deployment of these models, incorporating dynamic biomarker monitoring and immunomicroenvironment profiling, to achieve multidimensional, sustainable precision medicine for every NMIBC patient.

Despite these promising findings, several limitations should be acknowledged. First, most available studies are retrospective, single-center, and based on relatively small or highly selected cohorts, limiting the generalizability of AI models across diverse clinical settings. Additionally, only a limited number of studies included external validation of their results, which remains essential to ensure the robustness, reproducibility, and clinical applicability of AI models across different populations and institutions. Second, heterogeneity in data acquisition, imaging protocols, histopathological processing, and annotation standards introduces variability that may affect model robustness and reproducibility. Third, the narrative design of this review, combined with the absence of standardized reporting guidelines and formal quality assessment tools across included studies, as well as the restriction to English-language studies, limits the robustness of cross-study comparisons and increases the risk of bias in the evidence synthesis. Fourth, many AI systems function as “black boxes”, raising concerns regarding interpretability, clinical trust, and regulatory approval. Finally, the integration of multimodal data remains incomplete, and the lack of standardized reporting and benchmarking frameworks further complicates comparison between studies. Addressing these challenges will be essential to ensure safe, reliable, and equitable implementation of AI in NMIBC management.

Conclusions

AI improves the management of NMIBC by providing enhanced precision in diagnosis, risk stratification, and treatment planning. AI-assisted cystoscopy, advanced imaging, and computational histopathology deliver accurate, reproducible, and real-time analyses, reducing inter-observer variability. Multimodal AI models that integrate clinical, histopathological, and molecular data show promising results in predicting recurrence, progression, and treatment response, enabling personalized surveillance and therapeutic strategies that outperform current scoring systems. Despite practical and ethical challenges, AI represents a significant step toward more dynamic and individualized NMIBC care, with the potential to improve outcomes and inform current standards.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.