Lost in Tracking: Uncertainty-guided Cardiac Cine MRI Segmentation at Right Ventricle Base

TL;DR

This work complemented the public resource by reannotating the RV base in the ACDC dataset, with refined delineation of the right ventricle outflow tract (RVOT), and proposed a novel dual encoder U-Net architecture that leverages temporal incoherence to inform the segmentation when interplanar motions occur.

Abstract

Accurate biventricular segmentation of cardiac magnetic resonance (CMR) cine images is essential for the clinical evaluation of heart function. However, compared to left ventricle (LV), right ventricle (RV) segmentation is still more challenging and less reproducible. Degenerate performance frequently occurs at the RV base, where the in-plane anatomical structures are complex (with atria, valve, and aorta) and vary due to the strong interplanar motion. In this work, we propose to address the currently unsolved issues in CMR segmentation, specifically at the RV base, with two strategies: first, we complemented the public resource by reannotating the RV base in the ACDC dataset, with refined delineation of the right ventricle outflow tract (RVOT), under the guidance of an expert cardiologist. Second, we proposed a novel dual encoder U-Net architecture that leverages temporal incoherence to inform the segmentation when interplanar motions occur. The inter-planar motion is characterized by loss-of-tracking, via Bayesian uncertainty of a motion-tracking model. Our experiments showed that our method significantly improved RV base segmentation taking into account temporal incoherence. Furthermore, we investigated the reproducibility of deep learning-based segmentation and showed that the combination of consistent annotation and loss of tracking could enhance the reproducibility of RV segmentation, potentially facilitating a large number of clinical studies focusing on RV.

Figures (5)

• Figure 1: (a) Anatomy of LV and RV. The basal imaging plane covers the right ventricle outflow tract (RVOT), pulmonary valve (P), and tricuspid valve (T) sheehan2008right. (b) A short-axis basal slice contains atria (in green), P (in blue), and RVOT (in red), with complex and varying layouts. (c) Motion tracking has high uncertainty here ($u_b$ and $u_s$, defined in Section \ref{['sec:loss-of-tracking']}), indicating loss-of-tracking. (d) RV segmentation by 10 Bayesian ensembles exhibits high uncertainty, resulting in a poorly reproducible volume estimation ranging from 0.2 to 37.8 mL.
• Figure 2: The Dual-Encoder UNet architecture for CMR segmentation: the upper path encodes the original image, while the lower path encodes the "loss-of-tracking" from $I_t$ to $I_{t+\delta t}$, identified by a Bayesian motion-tracking model. The decoder $D$ predicts the segmentation map from the fused image feature $f_I$ and loss-of-tracking feature $f_{\phi}$.
• Figure 3: Qualitative results of tracking uncertainty and segmentation. The left panel shows the tracking uncertainty $u_b$ and $u_s$ between $I_t$ and $I_{t+\delta t}$. The right panel shows the segmentation labels and predictions. In case (b), the atrium and valve (cyan arrows) coexist with RV, and should not be included (c.f. the anatomy in Fig. \ref{['fig:teaser']}).
• Figure 4: Distribution of segmentation reproducibility as measured by volume standard deviation $\sigma_v$. Statistics (mean ± std) are given in corresponding colors.
• Figure 5: Examples of RVOT segmentation and reproducibility. High uncertainty indicates strong disagreement among different ensemble models. (a) Models trained with the original labels are uncertain on basal slices with both the valve and atria (cyan arrows) inplane. (b) The reproducibility is largely improved by the new annotations, and further reduced by the proposed method.