Deep learning diagnostics for bladder tumor identification and grade prediction using RGB method

Abstract

We evaluate the diagnostic performance of deep learning artificial intelligence (AI) for bladder cancer, which used white-light images (WLIs) and narrow-band images, and tumor grade prediction of AI based on tumor color using the red/green/blue (RGB) method. This retrospective study analyzed 10,991 cystoscopic images of suspicious bladder tumors using a mask region-based convolutional neural network with a ResNeXt-101-32 × 8d-FPN backbone. The diagnostic performance of AI was evaluated by calculating sensitivity, specificity, and diagnostic accuracy, and its ability to detect cancers was investigated using the dice score coefficient (DSC). Using the support vector machine model, we analyzed differences in tumor colors according to tumor grade using the RGB method. The sensitivity, specificity, diagnostic accuracy and DSC of AI were 95.0%, 93.7%, 94.1% and 74.7%. In WLIs, there were differences in red and blue values according to tumor grade (p < 0.001). According to the average RGB value, the performance was ≥ 98% for the diagnosis of benign vs. low-and high-grade tumors using WLIs and > 90% for the diagnosis of chronic non-specific inflammation vs. carcinoma in situ using WLIs. The diagnostic performance of the AI-assisted diagnosis was of high quality, and the AI could distinguish the tumor grade based on tumor color.

Overview of current applications and trends in artificial intelligence for cystoscopy and transurethral resection of bladder tumours

“…Shkolyar et al examined only CIS and reported at the AUA2023 [15]. In addition to white light imaging, narrow band imaging, Photodynamic diagnosis using blue light imaging has been reported [7,15,16]. Yoo et al focused on the three primary colours of light: red, green, and blue [7].…”

Sequential modeling for cystoscopic image classification

“…Whereas other studies evaluated their models on a select set of frames, our validation set captures the entirety of the diagnostic portion of TURBT videos. While prior studies focused on either the detection of cancerous bladder tumors or the detection of specific bladder tumor morphologies [9,10,29], we developed and evaluated a series of sequential models on a dataset containing bladder cancer of varying stages, grades, and appearances. We also included pathology-confirmed benign lesions in the dataset.…”

A Comparative Study of Deep Neural Networks for Real-Time Semantic Segmentation during the Transurethral Resection of Bladder Tumors

“…The 92.91% Dice score achieved by our PAN model indicates better tumor localization ability than reported in previous studies. As listed in Section 1 , the attention mechanism-based U-Net of Zhang et al [ 19 ] achieved a Dice score of 82.5%, the mask region-based CNN of Yoo et al [ 6 ] obtained a score of 74.7%, while the CystoNet of Shkolyar et al [ 4 ] did not report the Dice score result, but their per-frame sensitivity of 90.9% is lower than the 94.79% of our PAN model. This improvement in accuracy may be the result of more appropriate pre-processing of the input image, a better choice of neural network architecture, or a difference in the dataset.…”

“…Deep learning (DL) methods are developing rapidly, and as a result, neural network-based bladder cancer diagnosis is becoming increasingly popular [ 4 , 5 , 6 ]. DL solutions can not only achieve diagnostic accuracy similar to that of experienced specialists [ 2 , 7 ], but their results are also objective and reproducible [ 3 ], as they are based on mathematical operations.…”

“…Lastly, Yoo et al [ 6 ] deployed a mask region-based convolutional neural network with a ResNeXt-101-32×8d-FPN backbone to identify and grade bladder tumors on white-light and narrow-band cystoscopy images. The sensitivity and specificity they achieved were 95.0% and 93.7%, respectively, but the Dice score of localization was only 74.7%.…”