Deep Learning Analysis of Indocyanine Green Fluoroscopy of Ureters in Robotic Cystectomy: Toward Reducing Ureteroenteric Strictures

Abstract

Objectives:

To develop a deep learning method to quantify ureter perfusion during indocyanine green (ICG) fluoroscopy in robot-assisted radical cystectomy (RARC).

Introduction:

Ureteroenteric stricture (UES) is a common and clinically significant complication of RARC. Ureter ICG fluoroscopy likely reduces UES rates. However, its current reliance on subjective visual interpretation warrants a quantitative assessment method.

Materials and Methods:

A single-center retrospective pilot study was performed using 96 videos of RARC with ICG fluoroscopy, recorded between November 2019 and June 2024. Afterwards, 251 full-color and 358 ICG frames of suspended ureters were randomly split into 80/20% training and test sets and were annotated. A surgical foundation model, SurgeNetXL, which was readily pre-trained, was trained on the training sets. Performance was assessed using the Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD). Each ureter segmentation was divided into 50 horizontal planes with green intensity sampling points to calculate five perfusion parameters, including time to peak and normalized peak slope. These parameters were combined into a single composite “perfusion score” per plane. A stable and manually annotated video was used for a proof of concept.

Results:

The model performed well in segmenting full-color ureters (DSC = 0.800, HD = 24 pixels). Regarding ureters in ICG fluoroscopy, the model also performed well (DSC = 0.767, HD = 61 pixels), despite worse visibility. There was a good subjective visual correspondence between the perfusion score graphical overlay and the green intensity in the video.

Conclusions:

This pre-clinical study shows promising steps toward automated assistance of ureter dissection plane determination during RARC, aimed at reducing UES in a data-driven manner. Additional steps are required to improve the technique and to enable fully automated intensity measurements, which are needed to study correlations with UES outcomes.