DeepLumen Fast and Accurate Segmentation of Coronary Arteries for - PowerPoint PPT Presentation

DeepLumen Fast and Accurate Segmentation of Coronary Arteries for Improved Cardiovascular Care Kersten Petersen, Michiel Schaap, David Lesage, Matthew Lee, and Leo Grady GTC 2017

How do we find the right treatment for patients with symptoms of coronary artery disease (CAD) ? Image from Cardiac Health 2

Agenda Coronary Artery Disease HeartFlow Analysis DeepLumen

Coronary Artery Disease (CAD) • 1/3 of all global deaths are from CAD • 500k people in U.S. die of a fatal heart attack each year 2 • >$200B spent each year in the US alone on the diagnosis and treatment of patients with CAD 2 1. Atlas of Heart Disease and Stroke, WHO, 2004 2. Heart Disease and Stroke Statistics--2011 Update : A Report From the American Heart Association, Circulation 2011, 123:e18-e209 4

Coronary Artery Disease (CAD) Plaque (calcium, fat, cholesterol, fibrin, cellular waste) Lumen (inner part of artery) Plaque in the artery walls can obstruct the blood flow to the heart. Images from Texas Heart and Lakeland Health 5

Treatment Options Obstructive Disease Non-obstructive Disease Lifestyle Changes Medication Stent Bypass Images from news.com.au and Wikipedia 6

Usual Clinical Pathway Non-invasive Tests Lifestyle Changes or Patient with Medication non-obstructive disease Invasive Cath Lab Patient with Stent or Bypass obstructive disease 7

THE PROBLEM Currently, many patients are unnecessarily sent to the Cath Lab, where they face an expensive , time-consuming , and invasive test. Image from MedStar Franklin Square Medical 8

Agenda Coronary Artery Disease HeartFlow Analysis DeepLumen

Agenda Coronary Artery Disease HeartFlow Analysis DeepLumen

HEARTFLOW ANALYSIS A non-invasive CAD test that is more accurate than existing non-invasive CAD tests. 11

HeartFlow’s Clinical Pathway HeartFlow Analysis Lifestyle Changes or Patient with Medication non-obstructive disease far fewer Invasive Cath Lab Patient with Stent or Bypass obstructive disease 12

IDEA Calculate fractional flow reserve. 13

Fractional Flow Reserve (FFR) • Most precise CAD test (gold standard) for making treatment decision. • However, FFR is invasive , expensive, and time-consuming. Pd FFR = Pa Distal Pressure (Pd) Proximal Pressure 1. De Bruyne et al., NEJM 201 2. Pills et al., JACC 2007 (Pa) 3. Tonino et al., NEJM 2009 14

HeartFlow Analysis CT data submitted Anatomic model Physiologic model Functional assessment with HeartFlow Analysis delivered Computational Fluid Dynamics 15

Comparison to FFR (Gold Standard) 1 0.9 HeartFlow Analysis 0.8 Specificity 0.7 0.6 All other existing CAD Tests 0.5 0.4 1. Koo et al., JACC 2011 0.3 2. Min et al., JAMA 2012 3. Norgaard et al., JACC 2014 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 4. Norgaard et., Eur Radiol 2015 Sensitivity 16

PLATFORM Trial • Prospective multi-center clinical trial HeartFlow Analysis No clinically adverse events 1. Douglas et al., EHJ 2015 2. Hlatky et al., JACC 2015 • Cost savings of 26% to the health system after 3. Douglas et al., JACC 2016 accounting for a $1,500 cost of the HeartFlow Analysis 17

The HeartFlow Analysis • CE Mark (07/2011) • >150 peer-reviewed • HeartFlow announces publications collaboration with • De novo 510(k) FDA Siemens Healthineers clearance (11/2014) • >150 issued and to offer an integrated allowed patents • Regulatory solution for worldwide approval in Japan noninvasive (11/2016) assessment of CAD • NICE Guidance (announced 03/2017) recommends HeartFlow Analysis (02/2017) 18

Commercial Use Over 10,000 patients have received the HeartFlow Analysis worldwide 19

Agenda Coronary Artery Disease HeartFlow Analysis DeepLumen

Agenda Coronary Artery Disease HeartFlow Analysis DeepLumen

The Analysis Process CT data submitted Anatomic model Physiologic model Functional assessment with HeartFlow Analysis delivered Computational Fluid Dynamics 22

CT Data to Anatomic Model CT data submitted Anatomic Model Heart & Large Structures Vessel Paths Lumen 23

Quality Control Manual correction by 
 certified experts Heart & Large Structures Vessel Paths Lumen Anatomic Model 24

Challenges Top : Lumen CT data, Bottom : (Pixel) annotations of lumen 25

Calculating Lumen: Input CT Data Vessel Paths 26

Calculating Lumen: Curved Planar Representation (CPR) Vessel Paths Single Vessel (CPR) Single Vessel (CT volume) 27

Calculating Lumen: Curved Planar Representation (CPR) Frame Optimization Vessel Paths Single Vessel (CPR) Single Vessel (CT volume) 28

Calculating Lumen: Curved Planar Representation (CPR) GPU accelerated 
 PREDICTION Image (CPR) Landmarks (CPR) 29

Calculating Lumen: GPU accelerated 
 SURFACE RECONSTRUCTION Mesh Output Landmarks (CPR) Landmarks (CT volume) Mesh (CT volume) 30

The Focus of this Talk Image (CPR) Landmarks (CPR) 31

Pixel Classification? No! PROBLEMS 1. Can produce spurious components. 2. Can produce holes in segmentation. 3. No sub-voxel accuracy 32

Regression Regress distances from vessel path point (red) to vessel boundary at fixed angles . 33

Rotational Symmetry Frame 34

Ring Representation 35

Ring Representation 36

Ring Representation 37

Unfold the Ring 1) Concatenate frames. 2) Predict distance from vessel path (red) to upper vessel boundary. 38

Unfold the Ring , … , , … Apply the same model rotationally to predict one distance at a time. 39

Cyclic Padding r Cyclic Padding 2r 40

Cyclic Padding r Cyclic Padding 2r 41

Extension to 3D Frames of CPR Longitudinal Slice of CPR 42

3D Ring Representation 43

3D Ring Representation h w 2r Concatenate frame to padded unfolded 3D ring 44

DeepLumen Prediction r distance predictions 3D regression CNN 3D feature map: w * h * 2r (padded unfolded 3D ring) 45

Speed • Given the vessel line and the image (512 3 voxels), Deep Lumen takes 8 seconds to segment all coronary arteries using a GeForce GTX Titan X . 46

Comparison to State-of-the-Art 3D CNN DeepLumen 3D U-Net Ronneberger, et al. MICCAI 2016 47

Comparison to Ground Truth DeepLumen Ground Truth 48

Geometric Error • 1500 training images and 1500 testing images • We evaluated the median shortest distance between the mesh from automated DeepLumen and the ground truth mesh. • Healthy regions: 0.08 mm 
 Diseased regions: 0.10 mm • Resolution of CT: ≈ 0.40 mm 49

Validating Minimal Lumen Area • IVUS: Intravascular Ultrasound (Resolution: ≈ 0.10 mm, invasive) • OCT: Optical Coherence Tomography (Resolution: ≈ 0.02 mm, invasive) 50

Comparison of Minimal Lumen Area (MLA) Study Comparison Expert / Auto r Leber et al. 2005 CT vs IVUS Expert 0.54 Caussin et al. 2006 CT vs IVUS Expert 0.88 Voros et al. 2011 CT vs IVUS Expert 0.65 Boogers et al. 2012 CT vs IVUS Expert 0.75 De Graaf et al. 2013 CT vs IVUS Automated 0.84 Doh et al. 2014 CT vs IVUS Expert 0.53 Park et al. 2015 CT vs IVUS Expert 0.89 Non-expert 0.82 radiologist Automated 0.80 DeepLumen (Ours) CT vs OCT Automated 0.91 (95% CI: 0.89-0.94) 51

Conclusion • DeepLumen is a core component of the HeartFlow Analysis. It segments vessels faster and more accurate than state-of-the-art 3D CNN architectures . • It is still important that human experts provide quality control by ensuring that the segmentations are accurate. • Serious clinical problems can be addressed by combining image analysis, deep learning and simulation/modelling . • All of these factors are essential to find the best treatment for a patient with symptoms of coronary artery disease. 52

Thank you 53

Design Choices • Dilated convolution • No pooling • ReLu activations • Batch normalization 54

Rotational Symmetry - a a Predict distance at 0º on = Predict distance at angle a image rotated by angle -a. 55