MP65-02 DIAGNOSTIC CLASSIFICATION OF CYSTOSCOPIC IMAGES USING DEEP CONVOLUTIONAL NEURAL NETWORKстатья из журнала
Аннотация: You have accessJournal of UrologyBladder Cancer: Basic Research & Pathophysiology III1 Apr 2018MP65-02 DIAGNOSTIC CLASSIFICATION OF CYSTOSCOPIC IMAGES USING DEEP CONVOLUTIONAL NEURAL NETWORK Okyaz Eminaga, Axel Semjonow, and Bernhard Breil Okyaz EminagaOkyaz Eminaga More articles by this author , Axel SemjonowAxel Semjonow More articles by this author , and Bernhard BreilBernhard Breil More articles by this author View All Author Informationhttps://doi.org/10.1016/j.juro.2018.02.2068AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookTwitterLinked InEmail INTRODUCTION AND OBJECTIVES The current study aims to evaluate the application of deep learning for the diagnostic classification of cystoscopic images obtained from the cystoscopic examination, which is indicated during the urologic evaluations. METHODS Four-hundred-seventy-nine cases representing 44 urologic findings were considered for our pilot study. The image color was linearly normalized and equalized by applying the contrast Limited adaptive histogram equalization. Because these findings can be viewed from every possible angles and side by cystoscopy, we ultimately generated images rotated by 10 grades and flipped vertically or horizontally, thereby resulting in 18,681 images. After image preprocessing, we developed different deep convolutional neural network (CNN) models (i.e., "ResNet50", "VGG-19", "VGG-16", "InceptionV3", and "Xception") and evaluated these models by F1-score. Further, we proposed two CNN concepts (i.e., 90%-filter size of the previous layer and harmonic-series filter size). A training set (60%), a validation set (10%) and a test set (30%) were randomly generated from the study dataset. All models were trained on the training set, validated on the validation set and evaluated on the test set. RESULTS The XCeption-based model achieved the highest F1-score of 99.52% and followed by models based on ResNet50 and the harmonic-series concept, which showed F1-scores of 99.48% and 99.45%, respectively. These models could identify all cystoscopic images with bladder cancer correctly. When we focus on the images misclassified by the model with the best performance, 7.86% of images showing bladder stones with indwelling catheter, and 1.43% of Bladder diverticulum were falsely classified. CONCLUSIONS The results of the current study underline the potential of deep learning in diagnostic classification for cystoscopic images. Our future work will focus on integrating AI-aided cystoscopy into clinical routine. © 2018FiguresReferencesRelatedDetails Volume 199Issue 4SApril 2018Page: e859 Advertisement Copyright & Permissions© 2018MetricsAuthor Information Okyaz Eminaga More articles by this author Axel Semjonow More articles by this author Bernhard Breil More articles by this author Expand All Advertisement Advertisement PDF downloadLoading ...
Год издания: 2018
Авторы: Okyaz Eminağa, Axel Semjonow, Bernhard Breil
Издательство: Lippincott Williams & Wilkins
Источник: The Journal of Urology
Ключевые слова: Bladder and Urothelial Cancer Treatments
Открытый доступ: bronze
Том: 199
Выпуск: 4S