(UroToday.com) The 2023 European Association of Urology (EAU) annual congress held in Milan, Italy between March 10th and 13th, 2023 was host to an abstract session of studies addressing non-muscle invasive bladder cancer, from the initial diagnostic setting to follow-up. Dr. Atsushi Ikeda presented the results of a study evaluating whether cystoscopy with artificial intelligence (AI) assistance can overcome differences between cystoscope products.
While cystoscopy remains essential for the diagnosis and treatment of bladder cancer, its diagnostic accuracy remains highly dependent on the experience and knowledge of the performing clinician. The study investigators developed a real-time diagnosis support system for cystoscopy examination based on artificial intelligence that can be connected to an existing cystoscopy system with a single cable.
Using the ResNet50 model as the deep learning model platform, the training dataset for this system comprised 2,102 white light still images (normal: 1,671; tumors: 431) obtained using an Olympus VHA cystoscope. The investigators applied the machine learning model to images obtained using a Karl Storz cystoscope by performing a technique that allows for the creation of probability maps that indicate candidate lesions depicted by endoscopic images.
During the study, a cystoscope manufactured by Karl Storz (HR-VIEW) was used to evaluate the captured white-light still images (normal: 394, tumor: 80). Candidate lesions were compared with the correct data obtained by an expert urologist, the reference standard. The investigators evaluated whether their system could correctly classify the presence or absence of a tumor in images.
In this training dataset, 78 images were true positives, 249 true negatives, 145 false positives, and 2 false negatives. With the decision threshold of the AI model for lesion detection pre-set at 0.75, the sensitivity was 97.5% and the specificity was 63.2% for an overall accuracy of 0.69. Dr. Ikeda next showed a video demonstrating cystoscopy with and without AI assistance and noted the obvious over-detection present with the AI assistance, as demonstrated by an excess of heat maps.
The authors concluded that although over-detection was high using images obtained with an Olympus cystoscope as training data, this system demonstrated the possibility of providing examination support for the Karl Storz cystoscope with good lesion detection results. Additional training with small dataset from the Karl Storz cystoscope would be needed to improve the over-detection. The authors maintain hope that bladder cancer treatment will be improved using artificial intelligence in clinical practice.
Presented by: Atsushi Ikeda, Department of Urology, University of Tsukuba, Institute of Medicine, Tsukuba, Japan
Written by: Rashid K. Sayyid, MD, MSc – Society of Urologic Oncology (SUO) Clinical Fellow at The University of Toronto, @rksayyid on Twitter during the 2023 European Association of Urology (EAU) Annual Meeting, Milan, IT, Fri, Mar 10 – Mon, Mar 13, 2023.