MP58-06 AUTOMATED STAGING OF T1 BLADDER CANCER USING DIGITAL PATHOLOGIC H&E IMAGES: A DEEP LEARNING APPROACH

Abstract

You have accessJournal of UrologyBladder Cancer: Basic Research & Pathophysiology II1 Apr 2018MP58-06 AUTOMATED STAGING OF T1 BLADDER CANCER USING DIGITAL PATHOLOGIC H&E IMAGES: A DEEP LEARNING APPROACH Muhammad Khalid Khan Niazi, Thomas Tavolara, Vidya Arole, Anil Parwani, Cheryl Lee, and Metin Gurcan Muhammad Khalid Khan NiaziMuhammad Khalid Khan Niazi More articles by this author , Thomas TavolaraThomas Tavolara More articles by this author , Vidya AroleVidya Arole More articles by this author , Anil ParwaniAnil Parwani More articles by this author , Cheryl LeeCheryl Lee More articles by this author , and Metin GurcanMetin Gurcan More articles by this author View All Author Informationhttps://doi.org/10.1016/j.juro.2018.02.1838AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookTwitterLinked InEmail INTRODUCTION AND OBJECTIVES The American Joint Committee on Cancer defines T1 bladder cancer as the invasion of tumor cells into the lamina propria (LP), yet pathologists often struggle to confirm LP and/or muscularis mucosae invasion using H&E stains from bladder biopsies. Accurate reporting of the depth of LP invasion is important because of the worsened clinical outcomes observed in patients with deep invasion. We present an automated image analysis method that can recognize urothelium (U), LP, and muscularis propria (MP) from digital images of H&E-stained slides of bladder tissue. Its future clinical applications include automated risk stratification of T1 bladder cancer patients based on depth of LP invasion. METHODS Our method uses transfer learning in conjunction with convolutional neural networks (CNN) to identify different bladder layers from H&E images of bladder biopsies. Transfer learning enables CNN to equip computers with an ability to recognize and apply relevant knowledge from previous learning experiences when they encounter new tasks. Our proposed method is trained, cross-validated, and tested on 50,000 non-overlapping regions of size 64x64 pixels captured at 10x magnification. These regions are labeled for U, LP, MP, red blood cells (RBCs), inflammation (I), and cautery artifacts (C) by a GU pathologist. Each test image is processed by the CNN, producing six confidence heatmaps for each label. The MP and U maps are converted to binary masks, which represent the position of their respective bladder feature with ones and everything else with zeros. Depth of LP invasion is measured by determining the shortest distance from a given tumor nuclei in LP to the closest pixel with value 1 of the U binary mask. RESULTS The proposed method has 93.0% agreement with the pathologist in identification of bladder layers when tested on images of bladder cancer biopsies acquired from 86 patients. Additionally, given a set of tumor nuclei within LP, it precisely and accurately determines the depth of LP invasion, opening up the potential to sub stage T1 bladder cancer. When trained on samples from only three classes (U, LP, and MP), the method confuses I for U, RBCs for MP, and C for U, LP, and MP 90.0% of the times. This effect disappears in a six class method, thus a six class method is desirable. CONCLUSIONS Through accurate, automated segmentation of bladder layers from images of H&E biopsies and a given set of tumor nuclei, the depth of LP invasion can be quantified, allowing pathologists to make more informed, expedited, and accurate diagnoses of T1 bladder cancer. © 2018FiguresReferencesRelatedDetails Volume 199Issue 4SApril 2018Page: e775 Advertisement Copyright & Permissions© 2018MetricsAuthor Information Muhammad Khalid Khan Niazi More articles by this author Thomas Tavolara More articles by this author Vidya Arole More articles by this author Anil Parwani More articles by this author Cheryl Lee More articles by this author Metin Gurcan More articles by this author Expand All Advertisement Advertisement PDF downloadLoading ...