Making Object Detection Work in Medical Imaging Idan Bassuk - PowerPoint PPT Presentation

Making Object Detection Work in Medical Imaging Idan Bassuk

Segmentation Detection Classification

4

5

Mature technology 6

Mature, Accurate and Fast 7

Changing the world, one killer app at a time 8

Changing the world, one killer app at a time Visual Intelligence (Military and Business ) 9

Changing the world, one killer app at a time Robotics 10

Changing the world, one killer app at a time Medical Imaging is Exploding 11

Changing the world, one killer app at a time Medical Imaging 12

First - Understanding the Algorithms

What is a Convolutional Feature Extractor? 14

Convolutional Feature Extractor 15

Convolutional Feature Extractor – Plug and Play Person 97% 16

What is a classification head? C1 97% C2 2.9% C3 0.1% 17

How to plug in regression? x 12 px y -10 px w 42 px h 50 px 18

Classification accuracy Before Human Level Deep (Approximate) Learning 19

How to REALLY Plug in Regression? Sliding Window BG 90% Bounding Box Coordinates 20

How to REALLY Plug in Regression? Sliding Window BG 93% Bounding Box Coordinates 21

How to REALLY Plug in Regression? Sliding Window Person 97% Bounding Box Coordinates 22

How to REALLY Plug in Regression? Sliding Window Unicycle Wheel 95% Bounding Box Coordinates 23

The Single-Stage Detection Algorithm Unicycle Wheel 95% Bounding Box Coordinates x=5, y=-2, w=62,h=66 24 * SSD, YOLO 9000

Guidelines for Adapting Object Detection to Medical Imaging

#1: Question the basic assumptions

#2: Being on top of the research Architecture Meta-Architecture Post Processing Detection without Focal Loss Soft NMS Pre-Training Deformable Multi-Task Learning Learned NMS Convolutions Deformable ROI-Cropping Feature Pyramid Networks

#2: Being on top of the research Solving Scarcity of Positive Data

#2: Being on top of the research Attention is Different

#3: Solving Detection in a Large 3D Volume Images of cats Cat scans (CT) 1024x1024x3 512x512x300 input size x30 times larger object size x10 times smaller 30

#4: Polyglot Data Patient Diagnosis Scan data 1. Demographics 2. Referral Letter 3. Past Scans & Reports 4. Scanner Meta-data 31

Key Takeaways

Key takeaways 1. Medical Imaging is one of the quickest growing bottlenecks in the medical world

Key takeaways 1. Medical Imaging is one of the quickest growing bottlenecks in the medical world 2. Object Detection is a mature technology with many promising applications in Medical Imaging

Key takeaways 1. Medical Imaging is one of the quickest growing bottlenecks in the medical world 2. Object Detection is a mature technology with many promising applications in Medical Imaging 3. Adaptation is the name of the game 4. It’s hard - but achievable

Thank you! Idan Bassuk idan@aidoc.com

#1: Being on top of the research The Basics - a. Faster R-CNN b. Single Shot Detector (SSD) c. R-FCN d. YOLO , YOLO 9,000 e. Speed/accuracy trade-offs for modern convolutional object detectors f. Tensorflow Object Detection API

#1: Being on top of the research Advanced Material: a. Fast R-CNN b. Deep Learning and Object Detection Tutorial by Ross Girshick and Kaiming He c. Focal Loss for Dense Object Detection d. Deep Residual Learning for Image Recognition e. Feature Pyramid Networks for Object Detection f. Beyond Skip Connections: Top-Down Modulation for Object Detection g. DSOD: Learning Deeply Supervised Object Detectors from Scratch h. Instance-aware Semastntic Segmentation via Multi-task Network Cascades i. Fully Convolutional Inance-aware Semantic Segmentation j. Deformable Convolutional Networks k. Mask R-CNN

Credits for graphics 1) Idan Bassuk - idan@aidoc.com 2) TED.COM - Joseph Redmon: How a Computer Learns to Recognize objects Instantly 3) Satellite Image 4) Facebook.com 5) Cloud Factory 6) Real-Time Grasp Detection Using Convolutional Neural Networks 7) ImageNet Classification with Deep Convolutional Neural Networks 8) Visualizing and Understanding Convolutional Networks 9) http://www.asimovinstitute.org - The Neural Network Zoo 10) You Only Look Once: Unified, Real-Time Object Detection 11) Pokemon Go - vg247.com 12) Neurala