Improving realism in patient-specific abdominal Ultrasound simulation using CycleGANs

E. MORIS; L. LO VERCIO; I. LARRABIDE

Rennes

Congreso; Compuer Assisted Radiology and Surgery Congress; 2019

Purpose: Intravascular images (IVUS) allowsthe visualization of the lumen and atherosclerotic plaque.The automatic segmentation of IVUS series is a valu-able tool for computer-assisted diagnosis. However, thepresence of morphological structures difficults this task.This work explores the use of Convolutional Neural Net-works (CNNs) to automatize the classification of shad-ows, calcified plaques and bifurcations on IVUS images.Methods: CNN showed to be a great Deep Learning ar-chitecture for image classification. Four different tradi-tional CNN architectures, with different hyper-parameterssettings, were tested in the classification of morphologi-cal structures in IVUS images and results between themwere compared.Results: LeNet rendered a F score M of 0.52, while dif-ferent versions of the CNN proposed by Lekadir et al.for plaque characterization had a F score M of 0.58 and0.29. The best results where produced by VGG19, witha F score M of 0.6, which is the pre-trained architecturewith less convolutional layers.Conclusion: LeNet and VGG19 produced good resultsin the detection of morphological structures in IVUSimages. Best results were obtained with simpler archi-tectures of the networks.