Experience-dependent diagnosis. Cystoscopy is the gold standard for bladder cancer detection, but the accuracy of diagnosis depends heavily on the experience of the performing urologist. Distinguishing malignant from benign lesions, identifying flat lesions like carcinoma in situ (CIS), and predicting tumor grade are all subjective assessments prone to variability.
Two specific diagnostic challenges are particularly consequential: differentiating CIS from chronic non-specific inflammation (CNI) -- which look similar visually but have vastly different treatment implications -- and distinguishing between low-grade and high-grade urothelial carcinomas, which determines whether deep muscle biopsy is needed.
Consequences of misdiagnosis at TURBT. Transurethral resection of bladder tumor (TURBT) is the primary surgical procedure for bladder cancer. If the tumor grade is misjudged, an insufficient resection may leave cancer behind. Guidelines require a second TURBT if muscle biopsy is not performed or if the initial TURBT is incomplete -- adding procedural burden that improved upfront diagnosis could prevent.
Enhanced imaging modalities such as narrow-band imaging (NBI) and blue-light cystoscopy were developed to overcome the limitations of standard white-light cystoscopy for detecting flat lesions and small tumors. However, even these advanced tools are insufficient in some scenarios, motivating the introduction of AI-based diagnostic assistance.
A large retrospective image dataset. The study analyzed cystoscopic images from 1,010 consecutive bladder cancer patients and 290 patients with benign findings, collected between January 2017 and December 2020. After quality screening and additional curation, 10,991 images were used for AI training and testing, representing one of the largest cystoscopy image datasets used for this purpose at the time of publication.
Images were captured using a standardized flexible cystoscope and included both white-light images (WLIs) and narrow-band images (NBIs). Histopathological confirmation was the reference standard for all labels. A urologist with over 10 years of TURBT experience manually annotated tumor margins and classified all lesions.
Mask R-CNN with ResNeXt backbone. The AI model (Robin-Cysto, Infinyx Corporation) used a Mask Region-based Convolutional Neural Network (Mask R-CNN) with a ResNeXt-101-32x8d-FPN backbone. This architecture was chosen because it performs three tasks simultaneously: classifying the image, drawing bounding boxes around regions of interest, and generating pixel-level masks of tumor contours.
The dataset was split into training (8,244 images), validation (1,847 images), and test (900 images: 600 cancer, 300 benign) sets. The model was trained over 90,000 steps with a batch size of 6, using stochastic gradient descent with momentum. Multiple loss functions were combined for classification, bounding box regression, and pixel-level mask prediction.
A novel innovation: quantifying tumor color. Beyond detecting and outlining tumors, the study introduced a novel approach to grade prediction: extracting the average RGB (red-green-blue) color values from within the tumor contours identified by the AI. This converts subjective visual color impressions into objective numerical data.
RGB values were calculated for each tumor image by averaging the pixel-by-pixel color values across all pixels within the AI-identified tumor boundary. This average RGB was then used as a feature for a separate machine learning model -- a Support Vector Machine (SVM) -- trained to classify tumor grades based on color alone.
Why color might encode grade information. The hypothesis underlying this approach is that bladder tumors of different grades have different visual characteristics visible under cystoscopy. High-grade tumors tend to be more vascularized, with potentially different surface texture and color compared to low-grade tumors. Carcinoma in situ, a flat high-grade lesion, has different optical properties than the inflammation it mimics.
Color analysis was performed separately for WLI and NBI images. The SVM classifier found the optimal separating hyperplane -- a mathematical boundary that maximally separates the color distributions of different tumor grade groups. RGB values were compared across five categories: benign, CNI, CIS, low-grade urothelial carcinoma, and high-grade urothelial carcinoma.
High diagnostic accuracy for cancer vs. benign. The AI-assisted diagnostic device achieved sensitivity of 95.0% (285/300 cancer cases correctly identified), specificity of 93.7% (562/600 benign cases correctly classified), and overall diagnostic accuracy of 94.1% (847/900 total cases), with an area under the ROC curve (AUC) of 0.974.
These results match or slightly exceed those of previous AI cystoscopy studies. A contemporary review reported combined sensitivity of 89.7% and specificity of 96.1% across prior studies, while the current system achieved 95.0% sensitivity with 93.7% specificity -- better balancing the two metrics.
Tumor contour identification. The AI's ability to precisely outline tumor boundaries was measured by the Dice Score Coefficient (DSC), which measures the spatial overlap between AI-generated and physician-drawn contours. The AI achieved a DSC of 74.7%, compared to 67% in the only prior comparable study, which also measured contour identification for different lesion types.
Accurate tumor contour identification is clinically meaningful because precise delineation of tumor margins supports complete resection during TURBT, reducing the risk of leaving residual tumor behind -- a key driver of bladder cancer recurrence.
Grade-dependent color differences found in WLI. Statistical analysis of RGB values across tumor grade categories found significant differences in red and blue channel values by tumor grade in white-light images (p less than 0.001 for both). No significant color differences were detected across tumor grades in NBI, suggesting that the full color spectrum of standard white light captures grade-related information that NBI's restricted wavelength range cannot.
For the most clinically important comparison -- benign versus low-and-high-grade tumors -- the RGB-based SVM model using WLI achieved exceptional performance: sensitivity of 99.3%, specificity of 98.0%, and accuracy of 99.2%, with an AUC of 0.992. This means the AI's color analysis nearly perfectly distinguished cancer from benign disease in white-light images.
CIS versus CNI: a critical distinction. For differentiating CIS (high-grade flat cancer) from CNI (chronic inflammation that mimics CIS), the RGB model using WLI achieved sensitivity of 98.0%, specificity of 91.2%, accuracy of 93.0%, and an AUC of 0.972. This outperformed previously reported NBI-based specificity of 76.8% for this challenging comparison.
The most difficult comparison -- low-grade versus high-grade cancers -- was less well-classified with AUC of 0.701 in WLI, though the SVM was still able to draw a separating hyperplane between the two groups. This limitation likely reflects genuine biological similarity between low and high-grade tumors in their surface color properties.
White-light outperformed narrow-band imaging. A counterintuitive finding was that AI using standard WLI consistently outperformed NBI for color-based grade prediction. This is the opposite of what clinical experience with enhanced cystoscopy systems suggests for human observers.
The authors propose an explanation based on the physics of narrow-band imaging: NBI works by restricting light to short wavelengths (415 nm blue and 540 nm green), which are preferentially absorbed by hemoglobin, making blood vessels more visible. But this wavelength restriction also limits the range of colors that can be captured, potentially suppressing the intrinsic color differences between tumor grades that exist in the full visible spectrum.
White light preserves full color information. Standard WLI captures the complete visible spectrum of tumor surface color, allowing differences in red, green, and blue channels to reflect underlying tissue composition. By extracting average RGB values from the full-spectrum white-light images, the AI can detect color patterns that the restricted NBI spectrum would miss or obscure.
This finding has practical implications: it suggests that the most informative imaging modality for AI-based grade prediction may be the standard white-light cystoscope already in use everywhere, without requiring investment in specialized NBI or blue-light equipment for this particular AI application.
Internal validation only. All performance metrics were derived from internal testing on a per-image basis at a single institution. External validation using images from different hospitals, different cystoscope models, and different patient populations has not yet been performed. This is the most important limitation and means the reported accuracy figures cannot yet be generalized to other clinical settings.
The per-image analysis approach may overestimate performance compared to a per-patient analysis. Multiple images from the same patient share characteristics, potentially inflating apparent accuracy when images from the same patient appear in both training and test sets, though the authors attempted to mitigate this.
Color normalization not applied. The study did not perform color normalization -- a preprocessing step that adjusts for differences in image color balance due to different lighting, camera settings, or cystoscope configurations. Without normalization, RGB values measured in one setting may not translate reliably to another setup, which could affect model generalizability.
The authors also note that urine turbidity and the distance between the cystoscope and the lesion were not controlled, both of which can affect apparent color. A planned multicenter prospective study will address these limitations and may also incorporate RGB color data directly into the AI's training process, potentially enabling real-time grade prediction during cystoscopy.
A foundation for AI-guided endoscopy. This study demonstrates that deep learning can simultaneously detect bladder cancer, outline tumor boundaries, and predict tumor grade from color information -- all from standard cystoscopic images. Each of these capabilities addresses a real unmet clinical need in bladder cancer management.
The ability to predict grade during cystoscopy -- before histopathological analysis is complete -- could guide real-time decisions during TURBT: how deeply to resect, whether muscle biopsy is needed, and how urgently to plan follow-up.
Toward real-time AI-assisted cystoscopy. The authors envision this system as the foundation for a real-time AI-assisted cystoscope that provides the urologist with immediate feedback during the procedure. Such a system could standardize diagnostic accuracy across experience levels, making high-quality cystoscopic evaluation accessible even in settings where expert urologists are not readily available.
As training datasets expand, external validation is completed, and color normalization and real-time processing are implemented, AI-assisted cystoscopy has the potential to meaningfully reduce the misdiagnosis, understaging, and incomplete resection rates that currently drive bladder cancer recurrence.