RU-Net: A refining segmentation network for 2D echocardiography

Abstract

In this work, we present a novel attention mechanism to refine the segmentation of the endocardium and epicardium in 2D echocardiography. A combination of two U-Nets is used to derive a region of interest in the image before the segmentation. By relying on parameterised sigmoids to perform thresholding operations, the full pipeline is trainable end-to-end. The Refining U-Net (RU-Net) architecture is evaluated on the CAMUS dataset, comprising 2000 annotated images from the apical 2 and 4 chamber views of 500 patients. Although geometrical scores are only marginally improved, the reduction in outlier predictions (from 20% to 16%) supports the interest of such approach.

LU-Net: A Multistage Attention Network to Improve the Robustness of Segmentation of Left Ventricular Structures in 2-D Echocardiography

“…2) RU-Net, recently introduced in [33] and built from two cascaded U-Net1. The epicardial mask predicted by the first network is dilated and multiplied with the input image to provide a contextualized image as input to the second network, with a total number of 4M parameters.…”

“…Based on this information, a cropping strategy with pre-defined fixed dimensions was then applied to generate new images centered on the LV cavity. Leclerc et al also introduced a contextualization mechanism based on the multiplication of a binary map surrounding the union of the LV and the myocardium (derived from a first segmentation network) with the input image in order to provide as input a pre-processed image without irrelevant information to a U-Net model that performs the segmentation of left ventricular structures [33]. Results show that this method allows for a reduction of outliers in terms of segmentation results (from 20% to 16%) but unfortunately without any improvement in overall accuracy.…”

“…• Geometric outlier: the set of segmentation attached to a patient is seen as a geometric outlier if at least one of its eight corresponding distance scores (i.e. d m and d H values at ED and ES for both apical two and four-chamber views) is out of the corresponding bounds defined from the inter-observer variability [36]; • Anatomical outlier: the set of segmentation attached to a patient is seen as an anatomical outlier if the simplicity and convexity [33] of the corresponding segmented contours are lower than the lowest values computed from expert annotations on 50 patients. These two metrics hold values between 0 and 1, and are maximized for a circle.…”

Shape Constraints in Deep Learning for Robust 2D Echocardiography Analysis

“…For segmentation results, apart from the most used geometrical metrics: Dice coefficient, Hausdorff distance (HD) and Mean Surface Distance (MSD), we also use two anatomical metrics: Convexity(Cx) and Simplicity(Sp) [12].…”

“…Simplicity(Sp) = 4π * Area(P ) P erimeter(P ) (10) Based on these metrics, we calculate the number of outliers for algorithm/model robustness evaluation. The outlier of segmentation prediction for CAMUS dataset is established from the inter-variability tests with the upper limit values for HD and MSD, and lower limit values for the simplicity and convexity [12]. A prediction mask is considered as a geometrical outlier if its HD > 3.5mm or M SD > 8.2mm at ED, if HD > 4mm or M SD > 8.8mm at ES.…”