(UroToday.com) Dr. Bryn Launer presented the application of an artificial intelligence-based computer vision as an alternative approach to predict whether detected upper tract urothelial carcinoma (UTUC) was pathologically low grade (LG) or high grade (HG) during flexible ureteroscopy (fURS). Traditionally, endoscopic biopsy of the tumor is used to classify the tumor grade; however, this method is constrained by instrumentation and lack of visibility. Dr. Launer’s research is of importance as appropriate treatment plans require accurate tumor grading.
Dr. Launer and her team used 22 fURS videos taken from 19 patients who were being treated for UTUC tumors that were graded as LG or HG by pathology. Individual frames from these 30 FPS videos were then used to develop a convoluted neural network for computer vision to predict tumor grade. 80% of the frames (N=20,323 from 14 videos) were dedicated towards training the convoluted neural network while the remaining 20% (N=4064 from 8 videos) were used to test this network. An area under the receiver operating curve (AUC-ROC) assessed the accuracy, sensitivity, and specificity of the convoluted neural network to predict frame classification with a threshold of 0.5.
LG and HG UTUC tumors were pathologically classified in an equal number of videos. The mean LG and HG tumor video times were similar with an overall mean video time of 16s±9. The accuracy, sensitivity, and specificity were 0.73, 0.72, and 0.73, respectively.
The image below illustrates the ROC curve of the computer vision model’s performance.
The image below depicts an example of frame classification compared between pathology and the computer vision model whereby the red outline signifies incorrect classification by the model.
This model performed fairly for all thresholds with an AUC-ROC of 0.74. This model had an 87% correct prediction whereby 7 out of the 8 videos were correctly classified. As such, the model demonstrates good accuracy and performs well in predicting LG and HG frame classification.
Dr. Launer concludes that their computer vision model is a feasible method of classifying tumor grade as LG or HG and plans to release this software as an open source to further validate its robustness.
Presented by: Bryn Launer, PGY3, M.D, Department of Urology, Vanderbilt University Medical Center, Nashville, TN , @BrynLauner on X during the 2024 American Urological Association (AUA) Annual Meeting, May 3 – May 6, 2024, San Antonio, Texas
Written by: Victor Pham, B.S., University of California Irvine, Irvine, CA, @victorpham01 on X during the 2024 American Urological Association (AUA) Annual Meeting, May 3 – May 6, 2024, San Antonio, Texas