Work

Pd36-02 Retroperitoneal Single Port Vs Transperitoneal Multiport Robot Assisted Partial Nephrectomy (rapn) in Patients With Highly Hostile Abdomens, a Tertiary Referral Center Comparative Analysis

Year: 2025
Type: article
Abstract: INTRODUCTION AND OBJECTIVE: Augmented reality (AR)could soon become an everyday surgical tool, superimposing additional information on the surgeon's vision.We are developing an AR software for robot-assisted partial nephrectomy (RAPN), with automated parenchyma recognition using artificial intelligence.This process requires live recognition of the surgical image by the AR system, which relies on previous segmentation of anatomical features for neural network training.We therefore aimed to assess the reliability of the segmentation step for neural network training.METHODS: We built up a database of 47,956 images from 51 NPRA videos.On each image, renal parenchyma, perinephric fat, and tumor were segmented by a group of trained non-medical annotators (Figure 1).We previously assessed surgeons' inter-annotator variability, with specificity ranging from 92.6% to 96.5% and Dice index from 80.8% to 88.1%.We therefore next assessed the error rate of non-physicians, by medical evaluation of a random set of segmented images depending on the phase of the surgical procedure.RESULTS: On a representative sample of 5,459 images from 9 NPRAs, segmentations by non-physicians were considered totally correct in 69.83% of the cases (Q1Q3: 52.01-78.39).Looking at the surgical phases (Table 1), we found a 95.3% (Æ5.96) success rate for renal dissection, 61.61% (Æ28.61) for tumor dissection, 64.66% (Æ31.60) for tumorectomy and 78.14% (Æ25.86) for the reconstruction phase.Dissection of the hilum, not systematically performed, appeared to be the most delicate, with only 48.32% (Æ38.84)success rate.Failure was generally linked to repeated misjudgment on many images.CONCLUSIONS: Our kidney segmentation process for neural network training has acceptable reliability, but there is room for improvement.Highlighting the phases that present the greatest difficulties for non-physician annotators will help improve our process and focus both review management and annotator training.
Language: English
Subfield: Surgery
Field: Medicine
Open Access status: bronze
OpenAlex ID: W4409268690
FWCI:
Cites: 1
Cited by:
Related to: 10