(UroToday.com) Artificial Intelligence (AI) was a popular topic at the first uroradiology moderated poster session of the 2022 AUA Annual Meeting. Katherine Fischer, MD presented research on training a convolutional neural network (CNN) – a type of AI framework typically used for imaging-based studies – to identify and highlight kidneys, ureters, and bladder on CT scans. This study was supported by NIH funding and in part by 2021-2022 Urology Care Foundation Research Scholar Award Program.
AI-enhanced automation of nephrolithiasis identification, measurements and outcome prediction from CT scans can be significantly aided by a fundamental step; algorithms capable of identifying the urinary collecting system and organs. In this study, Fischer et al. annotated the kidney, ureters and bladder slice by slice of 29 CT urograms, a process called segmentation. They trained a specific type of CNN called U-Net to then autonomously identify these structures and segment them. CNN algorithms “see” CT scans in terms of 3D pixels known as voxels which have a Hounsfield Units value assigned to them indicating of the radiolucency or radiopacity of particular region. Algorithms use these numerical values and patterns to “view” the CT scans. In this study, the authors employed methods like clipping the CT scan down to a particular spatial dimension (axial 2.5, sagittal 1.62; coronal 1.62) and restricting the algorithm to viewing only voxels of specific Hounsfield unit range (-200 to 450HU) to effectively reduce the amount of data their model needed to sort through while learning to identify the relevant anatomy. This was combined with specific configurations to the U-Net architecture including specific algorithmic measures to improve success. From the 29 CT urograms manually segmented for training, the CNN was trained on a randomly chosen 25 scans and tested on the remaining 4.
Figure 1: Automatically segmented kidneys, bladder and ureters by the author’s U-net CNN algorithm.
The algorithm was able to automatically segment the kidneys, ureters and bladder with some reasonable success. They evaluated the algorithm’s success using Dice score, a statistical analysis that mathematically takes the segmentations of the CNN and superimposes them on the manual segmentations and compares the degree of overlap. Here are their Dice scores by anatomical region: kidney 0.879 ± 0.058; ureter 0.610 ± 0.101; and bladder 0.644 ± 0.388.
With this moderate to high level of precision, the authors now seek to expand their data set to increase success of the model and improve its generalizability. She states that since this project they have improved the size of the data set, but a major limiter is the time intensive process of manually segmenting these scans (requiring up to 1 hour per scan!) They are also working to train the model combining contrast and non-contrast phases of the same scan to help the model eventually be able to identify these structures even on non-contrast scans.
Presented by: Katherine Fischer, MD Children’s Hospital of Philadelphia
Co Authors: Yuemeng Li, Benjamin Schurhamer, Justin Ziemba, Yong Fan, Gregory Tasian
Written by: Kalon L. Morgan, Incoming 4th-year Medical Student at Rocky Vista University College of Osteopathic Medicine, Leadership and Innovation Fellowship Training (LIFT) Scholar at Department of Urology, University of California, Irvine. @kalon_morgan on Twitter during the 2022 American Urological Association (AUA) Annual Meeting, New Orleans, LA, Fri, May 13 – Mon, May 16, 2022.