Learningbased image segmentation for IVUS images Raja Yalamanchili

Learning-based image segmentation for IVUS images Raja Yalamanchili Computational Biomedicine Lab 1

Intravascular Ultrasound (IVUS) imaging Figure credits: Montana State University http: //www. montana. edu/wwwai/imsd/diabetes/myocard. htm, Yale-New Haven Hospital. http: //www. ynhh-healthlibrary. org, Normatem. http: //www. normatem. com/vp. html 2

Anatomy of Blood Vessel 3

Problem Statement • Automatic segmentation of different layers of a vessel to study characteristics of plaques and vessels – Lumen/Intima border – Media/Adventia border 4

Significance • Manual segmentation of even one frame is time consuming • IVUS sequence consists of thousands of frames 5

Challenges: Low Contrast Adventia Media Lumen 6

Challenges: Image Appearance Images acquired with 20 MHz and 40 MHz catheter frequency 7

Challenges: Image Appearance (2) Same image with different transformation parameters 8

Challenges: Artifacts A. Ringdown artifact B. Guidewire artifact C. Acoustic Shadowing 9

Literature Review • Image-based methods – Sonka et al. , Birgelen et al. , Zhang et al. (intensity and gradient information combined with Computational methods ) – Haas et al. , Luo et al. , Hui-Zhu et al. , Cardinal et al. , dos Santos Filho et al. (texture, statistical, temporal properties of images) • RF-based methods – Nair et al. , Nasu et al. , Kawasaki et al. , O’ Malley et al. , Mendizabal-Ruiz et al. 10

Limitations • Image-based methods rely on image properties – Image appearance – artifacts • No way to correct the segmentation result • Difficult to create a training set that can include all variations 11

Active learning method Segmentation Algorithm User Interaction Preliminary Result Confidence Measure Update Segmentation Parameters Final Result 12

Thank you! 13