Mp56-02 Deep-Learning Approach to Renal Parenchyma Detection: A First Step Towards Augmented Reality to Guide Robot-Assisted Partial Nephrectomy (Rapn)

You have accessJournal of UrologyKidney Cancer: Localized: Surgical Therapy IV (MP56)1 May 2024MP56-02 DEEP-LEARNING APPROACH TO RENAL PARENCHYMA DETECTION: A FIRST STEP TOWARDS AUGMENTED REALITY TO GUIDE ROBOT-ASSISTED PARTIAL NEPHRECTOMY (RAPN) Abderrahmane Khaddad, Adrien Bartoli, Kilian Chandelon, Gaëlle Margue, Julie Desternes, Nicolas Bourdel, and Jean-Christophe Bernhard Abderrahmane KhaddadAbderrahmane Khaddad , Adrien BartoliAdrien Bartoli , Kilian ChandelonKilian Chandelon , Gaëlle MargueGaëlle Margue , Julie DesternesJulie Desternes , Nicolas BourdelNicolas Bourdel , and Jean-Christophe BernhardJean-Christophe Bernhard View All Author Informationhttps://doi.org/10.1097/01.JU.0001008940.44711.d4.02AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: RAPN is the standard treatment for localized kidney tumors. Imaging and 3D reconstruction enable the surgeon to prepare his procedure and guide him through the important steps of surgery. Augmented reality based on real-time image fusion aims to provide precise guidance during the surgical procedure. Image fusion requires recognition and localization of the visible renal parenchyma, and the Deep-Learning (DL) approach could solve this task. The objective was to evaluate the ability of an artificial neural network (ANN) to recognize renal parenchyma during a surgical procedure, after training it to. METHODS: We used a UNet neural network. Our training dataset consisted of 16,435 images, in which visible renal parenchyma was manually annotated, taken from 8 RAPN videos. Annotation rules were recorded in a Guidebook to make this task standardized and reproducible. The images used in the evaluation were classified into 5 groups of different complexity, regarding annotation complexity. The evaluation dataset comprised 454 images from 28 RAPN videos. Part of the evaluation dataset (100 images) was used to study intra- and inter-annotator variability. The whole dataset was used to compare ANN predictions with annotations by an expert human annotator. Dice overlay index and IoU (Intersection over Union) were used to compare annotations and predictions. RESULTS: The intra- and inter-annotator reproducibility study assessing annotation reproducibility revealed a Dice index between 92.40 and 93.83% and an IoU between 86.51 and 89.05%. Performance evaluation of the ANN revealed a Dice index of between 41.05 and 51.53% and an IoU of between 29.19 and 39.69%, depending on the image class. Average specificity was 97.98%, with 1.57% false positives, and average sensitivity 35.03%, with 13.72% false negatives. CONCLUSIONS: The ANN showed promising preliminary results, recognizing renal parenchyma surface with specificity, but with low sensitivity. The study showed significant intra- and inter-annotator variability when performing this complex task. This first step towards the development of an augmented reality surgical guidance system will soon enable the start of the clinical phase, although further enrichment of the training dataset is still required. Download PPT Source of Funding: None © 2024 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 211Issue 5SMay 2024Page: e925 Advertisement Copyright & Permissions© 2024 by American Urological Association Education and Research, Inc.Metrics Author Information Abderrahmane Khaddad More articles by this author Adrien Bartoli More articles by this author Kilian Chandelon More articles by this author Gaëlle Margue More articles by this author Julie Desternes More articles by this author Nicolas Bourdel More articles by this author Jean-Christophe Bernhard More articles by this author Expand All Advertisement PDF downloadLoading ...