AUA 2023: Deep Learning Algorithms for Multi-Region Bladder Cancer Segmentation

(UroToday.com) The 2023 American Urological Association Annual Meeting included an imaging and radiology session, during which Dr. Jazayeri gave an intriguing presentation regarding deep learning (DL) models that use convolutional neural networks (CNN). CNN employ advanced and quantitative image features inside MRIs to detect the occurrence of muscle invasion. Dr. Jazayeri and colleagues specifically sought to evaluate multiparametric magnetic resonance imaging (mpMRI) as an algorithm for the automatic detection of bladder tumors, given that prior work has demonstrated their feasibility for fully automated and high-accuracy image segmentation. Further assessment of these models is critical to ensure that their classification probabilities are on par with their correctness and to evolve the tools available for bladder cancer staging.

Forty-three patients with bladder masses were prospectively evaluated with a 3 Tesla pelvic mpMRI, which includes high resolution T2-WI, diffusion weighted imaging/apparent diffusion coefficient (DWI/ADC), and T1 weighted dynamic contrast enhanced sequences (DCE). Three DL models (MAnet, PSPNet, Unet++) were all tested with the four following learning objectives: cross-entropy (CE), CE+dice loss (DSC), CE+margin-based label smoothing (MBLS), and focal loss (FL). Evaluations were completed with multi-modal data that incorporated T2-WI, DWI/ADC, and T1-DCE. Finally, Dice similarities coefficient (DSC) and Hausdorff distance (HD) were employed to evaluate segmentation performance and expected calibration error (ECE) measured the calibration performance of each DL model.

Dr. Jazayeri concluded that the UNet model provided the best results among the three DL models. Furthermore, the models were able to achieve better segmentation overall on DW1/ADC and T1-W images compared to T2-WI images. He closed his presentation with the takeaway message that DL models have the ability to provide automatic bladder tumor segmentation on mpMRI images.

One particularly thought-provoking question posed by the moderator revolved around the penetration ability of the DL models and whether it could not only distinguish the presence of disease but also determine the depth of it. Although the study did not focus on this aspect, Dr. Jazayeri noted how establishing tumor segmentation is the first step to improving bladder cancer staging. Progressing the DL algorithm to be able to measure how deep the tumor is certainly planned as a future direction. Furthermore, the study only enrolled a small pool of patients; their next step would need to assess a larger sample for proper clinical application. As a final quote from Dr. Jazayeri, he expressed, “If any institutions want to send [our team] MRI images, we’d be more than happy to take them” as collaborative efforts would help in their clinical testing of DL.

Figure 1. The accuracy of different AI algorithms in tumor segmentation on T1-weighted image

Table 1. Characteristics of Different Deep Learning Models in Bladder Cancer Detection.

Presented by: Seyedbehzad Jazayeri, MD, Department of Urology, University of Florida, Jacksonville, FL

1. Julien Nicholas, Kazem Gumus, Seyed Behzad jazayeri, Deheeraj Reddy Gopireddy, Jose Dolz, Mark Bandyk. Deep Learning Algorithms for Multi-Region Bladder Cancer Segmentation [abstract]. In: American Urological Association Annual Meeting. April 28 – May 1, 2023, Chicago, Illinois