Cancer Metastases Classification in Histological Whole Slide Images

Slides:



Advertisements
Similar presentations
Patient information extraction in digitized X-ray imagery Hsien-Huang P. Wu Department of Electrical Engineering, National Yunlin University of Science.
Advertisements

Applications of one-class classification
Rich feature Hierarchies for Accurate object detection and semantic segmentation Ross Girshick, Jeff Donahue, Trevor Darrell, Jitandra Malik (UC Berkeley)
Lecture 6: Classification & Localization
Carolina Galleguillos, Brian McFee, Serge Belongie, Gert Lanckriet Computer Science and Engineering Department Electrical and Computer Engineering Department.
Automatic in vivo Microscopy Video Mining for Leukocytes * Chengcui Zhang, Wei-Bang Chen, Lin Yang, Xin Chen, John K. Johnstone.
1 Learning to Detect Objects in Images via a Sparse, Part-Based Representation S. Agarwal, A. Awan and D. Roth IEEE Transactions on Pattern Analysis and.
Automatic Face Recognition Using Color Based Segmentation and Intelligent Energy Detection Michael Padilla and Zihong Fan Group 16 EE368, Spring
Face Processing System Presented by: Harvest Jang Group meeting Fall 2002.
Face Recognition Using Neural Networks Presented By: Hadis Mohseni Leila Taghavi Atefeh Mirsafian.
Tal Mor  Create an automatic system that given an image of a room and a color, will color the room walls  Maintaining the original texture.
Hurieh Khalajzadeh Mohammad Mansouri Mohammad Teshnehlab
Fully Convolutional Networks for Semantic Segmentation
Deep Convolutional Nets
1 End-to-End Learning for Automatic Cell Phenotyping Paolo Emilio Barbano, Koray Kavukcuoglu, Marco Scoffier, Yann LeCun April 26, 2006.
Lecture 4a: Imagenet: Classification with Localization
Color Image Segmentation Mentor : Dr. Rajeev Srivastava Students: Achit Kumar Ojha Aseem Kumar Akshay Tyagi.
Rich feature hierarchies for accurate object detection and semantic segmentation 2014 IEEE Conference on Computer Vision and Pattern Recognition Ross Girshick,
Lecture 3b: CNN: Advanced Layers
Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition arXiv: v4 [cs.CV(CVPR)] 23 Apr 2015 Kaiming He, Xiangyu Zhang, Shaoqing.
Lecture 4b Data augmentation for CNN training
EE368 Final Project Spring 2003
Radboud University Medical Center, Nijmegen, Netherlands
Recent developments in object detection
CS 4501: Introduction to Computer Vision Object Localization, Detection, Semantic Segmentation Connelly Barnes Some slides from Fei-Fei Li / Andrej Karpathy.
Faster R-CNN – Concepts
CS 6501: 3D Reconstruction and Understanding Convolutional Neural Networks Connelly Barnes.
Summary of “Efficient Deep Learning for Stereo Matching”
Data Mining, Neural Network and Genetic Programming
[Ran Manor and Amir B.Geva] Yehu Sapir Outlines Review
Journal Club M Havaei, et al. Université de Sherbrooke, Canada
Computer Science and Engineering, Seoul National University
Convolutional Neural Fabrics by Shreyas Saxena, Jakob Verbeek
The Problem: Classification
Jure Zbontar, Yann LeCun
A Neural Approach to Blind Motion Deblurring
Article Review Todd Hricik.
Classifying Breast Cancer Stage
CAMELYON16 Challenge Matt Berseth, NLP Logix Jacksonville FL
Training Techniques for Deep Neural Networks
Deepak Kumar1, Chetan Kumar1, Ming Shao2
Efficient Deep Model for Monocular Road Segmentation
CS 698 | Current Topics in Data Science
CS 698 | Current Topics in Data Science
CS6890 Deep Learning Weizhen Cai
Brain Hemorrhage Detection and Classification Steps
R-CNN region By Ilia Iofedov 11/11/2018 BGU, DNN course 2016.
Project Implementation for ITCS4122
Scott Tan Boonping Lau Chun Hui Weng
A Convolutional Neural Network Cascade For Face Detection
Aoxiao Zhong Quanzheng Li Team HMS-MGH-CCDS
Introduction to Neural Networks
RGB-D Image for Scene Recognition by Jiaqi Guo
Cancer Metastasis Detection Using Local Features and Random Forests
Object Detection + Deep Learning
Object Detection Creation from Scratch Samsung R&D Institute Ukraine
Tuning CNN: Tips & Tricks
Object Tracking: Comparison of
Somi Jacob and Christian Bach
RCNN, Fast-RCNN, Faster-RCNN
Midterm Exam Closed book, notes, computer Similar to test 1 in format:
Midterm Exam Closed book, notes, computer Similar to test 1 in format:
Department of Computer Science Ben-Gurion University of the Negev
Deep Object Co-Segmentation
DRC with Deep Networks Tanmay Lagare, Arpit Jain, Luis Francisco,
VERY DEEP CONVOLUTIONAL NETWORKS FOR LARGE-SCALE IMAGE RECOGNITION
Multi-UAV to UAV Tracking
An introduction to Machine Learning (ML)
Shengcong Chen, Changxing Ding, Minfeng Liu 2018
Deep CNN for breast cancer histology Image Analysis
Presentation transcript:

Cancer Metastases Classification in Histological Whole Slide Images Farhad Ghazvinian Zanjani Sveta Zinger Peter H.N. de With Department of Electrical Engineering Eindhoven University of Technology Eindhoven, The Netherlands

Method Outline Tissue region segmentation Patch extraction Test-time color augmentation CNN CRF Blob analysis DBSCAN clustering Class label Input slide 2 Method

Tissue Region Extraction Ignoring the empty zero-filled regions Computing a threshold map Assigning an Otsu threshold to each pixel by sliding a window Clamping the computed threshold map by 10% of Global Otsu threshold Using Morphology for removing small isolated objects in the binary image 3 Method

Small isolated regions removal Tissue Region Extraction Small isolated regions removal Input Color Slide Threshold map Binary slide 4 Method

Random patch extraction from slides Data Sampling Random patch extraction from slides Cropped from original full resolution slides Select patches inside the tissue-region mask 256x256 pixels Balanced classes Ignore some marginal patches which contain both classes and less than 75% of dominant class 5 Method

Train-time Data Augmentation Flipped up/down/left/right Rotated 90/180/270 degrees (fast matrix computation) Color augmented Transforming to HSV color coordinates Adding random offset to H, S and V channels Brightness (V channel) has been scaled randomly On the fly implementation 6 Method

Convolutional Neural Networks Inception v3 vs GoogleNet (v1) Original inception module [1] Inception v3 module [1] ConvNet Model Accuracy on Camelyon16 GoogleNet [2] 98.4 % Inception v3 99.5 % [1] Szegedy, Christian, et al, 2016. [2] Winner of Camelyon 2016. https://camelyon16.grand-challenge.org/ 7 Method

Convolutional Neural Networks GoogleNet - Inception v3 Initial parameters have been optimized on ImageNet12 dataset Resizing input to 299x299 Changing the output softmax layer for two-class prediction Changing learning rate by monitoring the performance on validation set Weight decay regularization 8 Method

False positive bootstrapping False Positive Bootstrapping (hard-mining) Retraining the network after adding false positive prediction to the training set Effects: Increasing accuracy about 1% on test set Small reduction in recall ConvNet Model False positive bootstrapping Accuracy On Camelyon16 Inception v3 - 98.7 % 99.5 % 9 Method

Test-time Color Augmentation Two approaches can be devised for tackling with the high color variations of staining in pathology WSIs Normalizing all slides to the color space of a reference [1] Training color space variations to the model 10 [1] Bejnordi, B. Ehteshami, et al. 2016.

Histogram transformation Test-time Color Augmentation Using color deconvolution [1] for finding ROI of absorbing Hematoxylin and Eosin Using stain standardization method for color conversion [2] RGB  Hematoxylin  Eosin   DAB Deconvolution  Binarized  HSD color space Histogram transformation Template  Converted color  [1] Ruifrok, Arnout C., and Dennis A. Johnston, 2001. [2] Bejnordi, B. Ehteshami, et al. 2016. 11

Examples of Color Conversion (negative and positive patches) Hematoxylin  Eosin  Original converted Hematoxylin  Eosin  Original converted Center 1 Center 1 Center 2 Center 2 Center 3 Center 3 Center 4 Center 4 Center 5 Center 5 Negative samples Positive samples 12 Method

Color Conversion Between Samples of 5 Medical Centers Target Center 1 Center 2 Center 3 Center 4 Center 5 Center 1 Source Center 2 Center 3 Center 4 Center 5 13 Method

Test-time Color Augmentation Convert the color of test input patch to have a similar distribution to the training examples of different medical centers Select the less uncertain prediction of the network on the color- augmented set 14 Method

Post processing – Conditional Random Field (CRF) Zoom ROI GT PNet( x=tumor) > 0.5 PNet( x=tumor) > 0.9 CRF 15 Method

Blob Analysis and DBSCAN clustering Computing the major axes of a fitted ellipse to the tumor region Computing the area of Hematoxylin mask, inside the tumor region If slide has been labeled as itc or micro metastases, then check by DBSCAN clustering Input slide Detected positive regions clustered 16 Method

Result Our method on Camelyon17 test set shows 0.87 kappa score in patient-level 17 Result

Conclusion We used a machine learning method, based on convolutional neural networks Inception v3 has been used as a pixel classifier False positive bootstrapping improves the prediction performance Test-time color augmentation used for decreasing the prediction uncertainty Using Conditional Random Field improves the label assignment 18 Conclusion