Cystoscopy is the standard tool for bladder cancer diagnosis and surveillance, yet it has well-documented limitations that result in high misdiagnosis and incomplete resection rates.
White light cystoscopy is effective for papillary lesions but frequently misses flat cancerous tissue (carcinoma in situ) and very small tumors. The reported misdiagnosis rate under white light is approximately 30%.
The incomplete resection rate during transurethral resection of bladder tumors (TURBt) reaches up to 50%, contributing directly to high rates of early recurrence and disease progression.
Alternative technologies such as fluorescence cystoscopy and narrow-band imaging improve lesion detection but have seen limited adoption, highlighting the need for a more accessible and scalable solution.
A total of 69,204 cystoscopic images from 10,729 consecutive patients across six Chinese hospitals were collected to develop and validate the Cystoscopy Artificial Intelligence Diagnostic System (CAIDS).
Participating centers included both national-tier hospitals and municipal hospitals, ensuring representation of diverse patient populations, imaging equipment, and clinical environments.
Images were split into training and internal validation sets (80:20 ratio) from the two largest hospitals, while images from the remaining four hospitals served as independent external validation sets.
All diagnoses were histologically confirmed via biopsy, TURBt, or radical cystectomy, and patients were followed for at least 6 months to catch delayed diagnoses. Tumors were staged using the 2016 TNM system.
Twelve experienced urologists from the primary center participated in image quality assessment and labeling, with disagreements resolved by a professor with over 20 years of urology experience.
The CAIDS was built using a pyramid scene parsing network (PSPNet) with transfer learning from ImageNet. PSPNet was chosen after benchmarking against FCN, SegNet, DilatedNet, and CascadeNet architectures.
The preprocessing module standardizes images by center-cropping to square format, resizing to 256 by 256 pixels, and normalizing pixel values to between 0 and 1. Each training image was augmented into four versions including the original, vertical flip, horizontal flip, and slight rotation.
Feature extraction uses a ResNet-101 backbone pretrained on ImageNet, enabling the model to generalize from natural images to medical endoscopic images and helping mitigate the effects of class imbalance.
The pyramid pooling module processes feature maps at four different scales simultaneously, enabling the model to capture both large-scale global patterns and fine local details in bladder tissue, which is critical for identifying small or flat lesions.
The CAIDS achieved accuracy exceeding 97.7% across all validation sets, with near-ideal areas under the ROC curve ranging from 0.989 to 0.998.
Internal validation accuracy was 0.977, with sensitivity, specificity, and negative predictive value all exceeding 0.975. The positive predictive value was 0.819, with false positives primarily caused by bladder wall elevation and inflammation.
External validation accuracy reached 0.990 at SZSH, 0.982 at STCH, 0.978 at AMUFH, and 0.991 at NJFH, demonstrating strong generalization across centers with different equipment and patient demographics.
The precision-recall curves confirmed robust performance despite class imbalance, with areas under the PR curve between 0.949 and 0.986 across external centers. F1 scores across external sites ranged from 0.923 to 0.957.
An independent prospective validation set of 80 patients undergoing TURBt achieved sensitivity of 0.996, specificity of 0.770, and an AUC of 0.976, further supporting real-world clinical applicability.
A dedicated comparison subgroup of 260 images of complex lesions was assembled, including 130 carcinoma in situ or small tumor images and 130 chronic inflammation images, representing the most diagnostically challenging cases.
Sensitivity for the three groups of urologists was 0.508 for trainees, 0.562 for competent urologists, and 0.754 for experts -- compared to 0.954 for the CAIDS.
In terms of speed, trainees, competent urologists, and experts required an average of 45, 38, and 35 minutes respectively to evaluate the 260 images. The CAIDS completed the same task in just 12 seconds.
When urologists reviewed images after receiving CAIDS probability estimates, sensitivity improved dramatically: trainees rose from 0.508 to 0.731, competent urologists from 0.562 to 0.846, and experts from 0.754 to 0.908. This demonstrates that CAIDS functions effectively as a decision-support tool rather than a replacement.
A major advantage of CAIDS is its potential to equalize diagnostic quality across different hospital tiers and experience levels. TURBt performed by less experienced urologists is associated with higher early recurrence rates, and CAIDS can help bridge this gap.
For carcinoma in situ specifically, expert urologists misdiagnosed 24.6% of tumor lesions in complex cases. With CAIDS guidance, 62.5% of those misdiagnosed cases could be avoided, representing a substantial patient safety benefit.
The high negative predictive value of 0.989 to 0.999 across validation sets means that a negative CAIDS result reliably rules out bladder cancer, minimizing missed diagnoses without substantially increasing the rate of unnecessary biopsies.
The system has already been deployed in an operating room at Sun Yat-sen Memorial Hospital, processing live cystoscope video frames in real time with a latency under 50 milliseconds per frame using GPU acceleration.
The CAIDS is designed for seamless integration into existing cystoscopy workflows. It captures video streams from cystoscopes via standard DVI or S-Video cables without requiring specialized equipment.
Each video frame is converted to a normalized image, passed through the PSPNet model, and a lesion mask is predicted. The mask is overlaid on the original frame and displayed on a monitor in real time.
GPU acceleration enables processing speeds below 50 milliseconds per frame, ensuring that the system does not interrupt the procedure or distract the operating urologist.
The system's non-distracting real-time overlay format is an important design feature for clinical adoption, as any AI-assisted tool must integrate with rather than disrupt the urologist's natural workflow.
The CAIDS demonstrates that a large, multicenter, pathologically validated AI system can outperform experienced urologists on the most diagnostically challenging bladder cancer cases while being dramatically faster.
The system addresses two critical unmet needs: improving detection of commonly missed lesions such as carcinoma in situ and small tumors, and reducing procedure time and cognitive burden for the operating urologist.
Prospective clinical trials are the necessary next step to formally evaluate the CAIDS in routine practice, assess its impact on recurrence rates, and determine guidelines for its integration into standard cystoscopy protocols.
The authors note that generalization to non-Chinese populations remains to be validated, and safety considerations for widespread clinical deployment will need to be addressed before broader adoption.