A challenging disease to manage. Bladder cancer affects over 357,000 new patients globally each year and carries a recurrence rate of 50-80%, making it one of the most expensive cancers to manage over time. Decisions about treatment depend critically on accurate assessment of tumor grade and whether the cancer has invaded the bladder muscle.
The gold standard for bladder cancer detection is cystoscopy -- direct telescopic examination of the bladder. However, cystoscopy is invasive, expensive, and poorly sensitive for low-grade tumors (sensitivity of only 61%). The need to repeat it every 3-6 months for surveillance creates significant burden for patients.
Biopsy and staging errors are common. Tumor biopsy provides tissue diagnosis but cannot sample the entire tumor at any given moment, and intratumoral heterogeneity frequently causes misclassification of cancer grade or stage. This directly affects treatment selection, risking undertreatment of aggressive disease or overtreatment of indolent disease.
Beyond detection, predicting whether a patient will respond to BCG immunotherapy (for non-muscle-invasive disease) or neoadjuvant chemotherapy (for muscle-invasive disease) remains a major unsolved problem. Without reliable predictors, some patients receive ineffective treatments while missing the optimal window for alternative approaches.
What radiomics is. Radiomics, first described by Lambin in 2012, is based on the hypothesis that medical images contain far more quantitative information than radiologists typically extract through visual inspection -- including cellular and molecular-level data encoded in image texture and structure.
Rather than a radiologist describing an image subjectively, radiomics uses software to extract hundreds or thousands of quantitative features from imaging data -- including CT, MRI, PET, and ultrasound. These features describe tumor shape, intensity patterns, texture, and spatial heterogeneity in precise numerical terms.
Machine learning powers the analysis. Raw radiomic features are not clinically useful on their own. Machine learning (ML) algorithms are essential for selecting the most informative features and building prediction models that connect imaging patterns to clinical outcomes such as tumor grade, stage, treatment response, and survival.
The combination of radiomics and ML offers a non-invasive, repeatable, and low-cost way to capture information about the entire tumor volume -- including its spatial heterogeneity -- that no single biopsy can provide. This makes it a natural complement to precision medicine approaches in bladder cancer.
Step 1: Image acquisition. Radiomics begins with collecting standardized images. The most common formats (PACS and DICOM) are widely used across institutions, facilitating data sharing. However, variations in imaging protocols -- different scanner models, field strengths, contrast agents, and acquisition parameters -- can introduce inconsistencies that reduce reproducibility across studies.
Step 2: Volume of interest (VOI) segmentation. The tumor region must be precisely outlined before features can be extracted. This is the most labor-intensive step and the most prone to variability. Many tumors have indistinct borders, and different analysts may draw slightly different contours. Fully automated segmentation algorithms exist but all have limitations for complex or irregular tumor shapes.
Step 3: Feature extraction. Once the VOI is defined, software extracts features across four categories: (1) shape features such as volume, diameter, and sphericity; (2) first-order statistics such as mean, median, and entropy of voxel intensities; (3) texture features measuring spatial relationships between adjacent voxels (e.g., Gray-Level Co-Occurrence Matrix); and (4) high-order features derived from mathematical transforms that capture complex patterns across the image.
Step 4: Model building. Feature selection eliminates redundant and low-value features to prevent overfitting. An appropriate ML algorithm is then trained on the selected features to build a prediction model. Internal validation (e.g., cross-validation) and external validation (testing on an independent dataset) are both essential to confirm that the model generalizes beyond the training data.
Classifying tumor stage from CT. Garapati et al. developed an ML model using CT urography (CTU) data from 76 bladder cancer cases to classify tumors as Stage T2 or below/above. Morphological and texture features were both informative, and four different ML classifiers (LDA, neural network, SVM, random forest) all showed similar and promising staging accuracy, demonstrating that the imaging features -- not the specific ML algorithm -- are the primary driver of performance.
For tumor grading, Zhang et al. used diffusion-weighted MRI (DWI) and apparent diffusion coefficient (ADC) maps in 61 patients, selecting features via recursive feature elimination with SVM. The optimized classifier achieved an AUC of 0.861 with 82.9% accuracy, 78.4% sensitivity, and 87.1% specificity for distinguishing high-grade from low-grade bladder cancers.
Multiparametric MRI outperforms single-modality models. Wang et al. built five radiomic models from different MRI sequences (T2-weighted, DWI, ADC, and combinations) in 70 patients. The joint model combining all modalities achieved an AUC of 0.9276 -- significantly better than any single-modality model -- demonstrating the value of integrating complementary imaging information for preoperative grade prediction.
A key finding across these staging studies is that ML-based radiomic models can reduce bladder cancer staging error rates by up to 50% compared to conventional methods, offering a meaningful improvement that could directly affect treatment selection at the time of diagnosis.
Distinguishing muscle-invasive from non-muscle-invasive disease. The most clinically critical staging decision in bladder cancer is whether the tumor has invaded the bladder muscle wall, as this determines whether a patient needs radical cystectomy or can be managed more conservatively. Zheng et al. developed a radiomic-clinical nomogram using three-dimensional MRI features focused on the tumor base.
This nomogram achieved an AUC of 0.922 in the training cohort and 0.876 in an independent validation cohort for predicting muscle invasiveness -- performance that would be highly clinically useful for guiding treatment decisions without requiring surgical staging.
Predicting survival after radical cystectomy. Wang et al. compared seven ML algorithms for predicting 5-year mortality in 117 muscle-invasive bladder cancer patients who had undergone radical cystectomy. Regularized extreme learning machine (RELM) and extreme learning machine (ELM) models achieved the best results, with mean sensitivity and specificity both exceeding 0.8, and with the advantage of very fast computation speed.
For recurrence prediction, Xu et al. developed a multiparametric MRI-based radiomic nomogram incorporating clinical risk factors to predict 2-year recurrence risk. The combined radiomic-clinical model consistently outperformed either the radiomics or clinical model alone, highlighting the additive value of imaging information when combined with clinical data.
Chemotherapy response evaluation from CT. Cha et al. explored whether CT-based radiomics could predict chemotherapy response in bladder cancer, using a deep learning CNN for automated tumor segmentation. The computerized radiomic assessment improved AUC by 0.03 compared to manual contour-based analysis -- a modest but consistent improvement demonstrating the precision advantage of automated segmentation.
A second study by Cha et al. used three independent radiomic predictive models for chemotherapy response assessment, finding agreement between all three models and two expert radiologists. The researchers proposed the computer-aided system as a second-opinion tool that assists rather than replaces clinical judgment.
Neoadjuvant chemotherapy response in muscle-invasive disease. A CT-based decision-support system was evaluated in 123 patients for predicting complete or partial response to neoadjuvant chemotherapy. The combined AI-supported assessment (AUC 0.80) significantly outperformed physicians working without the tool (AUC 0.74, p less than 0.05), showing a statistically meaningful benefit from AI assistance.
These therapy response studies represent the most direct clinical application of radiomics in bladder cancer -- enabling oncologists to identify which patients are responding to treatment earlier, potentially allowing faster adjustments before toxicity accumulates from an ineffective regimen.
Small sample sizes and single-institution studies. The vast majority of published radiomics studies in bladder cancer use fewer than 200 patients from single institutions. When the number of extracted features greatly exceeds the patient count -- which is nearly always the case -- feature selection bias and overfitting become serious risks, potentially producing results that do not replicate in independent cohorts.
External validation is still the exception rather than the rule. A model that performs well on the data it was built from, but has not been tested on patients from a different institution or scanner, cannot yet be trusted for clinical deployment. This gap between research results and generalizable tools is the most important limitation of current radiomics work.
Lack of standardization across studies. Different research groups use different imaging protocols, segmentation methods, feature extraction software, and ML algorithms. This diversity makes it nearly impossible to directly compare results across studies or combine datasets for larger analyses. International consensus guidelines on radiomics methodology are needed but still absent.
Code and data transparency is also a significant issue -- most papers do not share the underlying code or image datasets, preventing independent verification of results. The field needs to move toward open-science standards where results can be reproduced and peer-checked before influencing clinical practice.
Connecting images to genomics. Radiogenomics -- the integration of imaging features with genomic data -- represents a promising frontier for bladder cancer. Lin et al. showed that a combined radiomics and transcriptomics model could predict progression-free interval with AUROCs above 0.94, suggesting that imaging features capture information about tumor biology at the molecular level.
This approach could eventually enable non-invasive tumor genomic profiling from imaging alone -- bypassing the need for biopsy and reducing the problem of sampling only a small part of a heterogeneous tumor. Several studies in other cancer types have already demonstrated that radiomics features can predict specific gene mutations, expression signatures, and immune phenotypes.
AI as a complement to clinical expertise. The authors emphasize that AI-supported radiomics should be viewed as a tool that augments clinical decision-making rather than replacing it. The combination of expert radiologist interpretation with quantitative AI analysis consistently outperforms either approach alone, and this collaborative model is the most realistic path toward clinical adoption.
For bladder cancer specifically, the full-cycle management potential of radiomics -- from initial detection through staging, treatment response monitoring, and recurrence surveillance -- makes it uniquely suited to transform a disease that currently requires frequent invasive monitoring into one that can be tracked non-invasively with high confidence.