Pd27-08 Pre-Training an AI Model Using Two Types of Automatically Generated Images for Cystoscopy AI Diagnosis of Bladder Cancer

You have accessJournal of UrologySurgical Technology & Simulation: Artificial Intelligence II (PD27)1 May 2024PD27-08 PRE-TRAINING AN AI MODEL USING TWO TYPES OF AUTOMATICALLY GENERATED IMAGES FOR CYSTOSCOPY AI DIAGNOSIS OF BLADDER CANCER Atsushi Ikeda, Ryuunosuke Kounosu, Wonjik Kim, Hirokazu Nosato, Yuu Nakajima, and Hiroyuki Nishiyama Atsushi IkedaAtsushi Ikeda , Ryuunosuke KounosuRyuunosuke Kounosu , Wonjik KimWonjik Kim , Hirokazu NosatoHirokazu Nosato , Yuu NakajimaYuu Nakajima , and Hiroyuki NishiyamaHiroyuki Nishiyama View All Author Informationhttps://doi.org/10.1097/01.JU.0001008580.58088.27.08AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: We have researched and developed a cystoscopy diagnostic support system using artificial intelligence (AI) to reduce lesions' oversight. While AI requires many images and labels, obtaining sufficient training data is not straightforward. Therefore, we have reported improved diagnostic accuracy through pre-training on cystoscopic images by utilizing transfer learning from the ImageNet dataset, a natural image dataset. However, natural images have mislabeling and copyright concerns, making them unsuitable for practical use. In this study, we evaluated the effectiveness of a formula-driven supervised learning (FDSL) approach to address these issues. We pre-trained two automatically generated images tailored to the characteristics of cystoscopy images and adapted them to the target task. METHODS: Using the ResNet50 AI model, we conducted a two-step learning process. First, we trained the model with two million automatically generated images, including one million fractal images representing texture features and one million visual atom images representing contour features. Subsequently, we performed learning using 8,812 cystoscopic images, where 1,259 contained tumor lesions, and 7,553 did not. We evaluated the AI model's performance on a test dataset of 422 cystoscopic images (87 with tumor lesions and 335 without), which were not used in training, and compared with the performance of urologists from previous studies. RESULTS: The AI model achieved an AUC (area under the curve) of 0.978, sensitivity of 0.933 at a Youden index of 0.908, specificity of 0.975, accuracy of 0.966, precision of 0.908, and an F1 score of 0.9199, to discriminate between images with tumor lesions and those without. CONCLUSIONS: The AI model, which underwent transfer learning on cystoscopic images, demonstrated superior performance to beginner urologists by leveraging FDSL with two automatically generated images for pre-training. In this study, we highlighted the effectiveness of this approach. We anticipated the development of a highly accurate cystoscopy diagnostic support model while reducing the cost of pre-training by harnessing FDSL with no problems that natural images have. Source of Funding: This study is based on results obtained from a project, JPNP20006, commissioned by the New Energy and Industrial Technology Development Organization (NEDO) © 2024 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 211Issue 5SMay 2024Page: e553 Advertisement Copyright & Permissions© 2024 by American Urological Association Education and Research, Inc.Metrics Author Information Atsushi Ikeda More articles by this author Ryuunosuke Kounosu More articles by this author Wonjik Kim More articles by this author Hirokazu Nosato More articles by this author Yuu Nakajima More articles by this author Hiroyuki Nishiyama More articles by this author Expand All Advertisement PDF downloadLoading ...