Whole-heart and Great Vessel Segmentation from 3D Cardiovascular Magnetic Resonance Images in Congenital Heart Disease (Part II)

Congenital heart disease (CHD) includes all heart defects existing at birth, encompassing a wide array of cardiac malformations and topological changes (see CHD.pdf at https://www.dropbox.com/sh/tn4hmfn31kmbllw/AACxPAXVaA1xF3pKHGf7LcZSa). These include: shape changes within a vessel or chamber, abnormal connectivity between cardiac structures, abnormal structure locations, duplicated structures, and/or missing structures. The heart of each CHD patient is different, with a unique combination of heart defects from birth, new atypical connections and implants from prior surgeries, and shape changes from long-term cardiac remodeling. Severe CHD requires multiple surgeries throughout life. For surgical planning, clinicians must understand each patient’s highly individual cardiac anatomy. Patient-specific, 3D surface models of the heart have great promise for surgical planning, whether virtual or 3D-printed. Studies have shown that they may lead to a greater appreciation of the true locations and sizes of intracardiac structures, aid decision making and consensus, and even change the surgical plan from that originally based on imaging! Cardiac MR is an attractive modality for preoperative imaging, from which heart models could be generated. MR captures cardiac anatomy at high-resolution while avoiding ionizing radiation (unlike CT - this is particularly important for children). However, segmentation is a prerequisite to create heart models for CHD surgical planning. The lack of accurate, robust segmentation tools that can handle the anatomical variability of CHD is a major bottleneck to widespread adoption of 3D heart models, since manual segmentation requires many hours per subject and is subject to inter- and intra-observer variability. We have previously organized the HVSMR challenge at MICCAI 2016 (http://segchd.csail.mit.edu, Zuluaga et al., 2016), before which very few image analysis methods for CHD had been developed. The challenge dataset was small, consisting of 20 3D cardiac MR images (10 training, 10 testing) to be segmented into the blood pool, surrounding myocardium, and background. Here, we propose HVSMR-II, with a larger, more challenging 3D cardiac MR dataset and a more difficult segmentation task. First, the challenge dataset has been increased to 90 scans, which include a much wider variety of very challenging heart defects. Second, we ask participants to perform “whole-heart segmentation”, by delineating all eight of the major cardiac structures (four chambers and four great vessels). Segmentation accuracy will be assessed by Dice scores (volume overlap score) and symmetric Hausdorff distances (maximum distance between the two segmentation surfaces). Clinical Impact - 3D heart models in which each structure is separately labeled may be more intuitive for surgical planning than blood pool models. Whole-heart segmentation would also facilitate automatic computation of several quantitative metrics of cardiac function in CHD patients, and also promises to enable future research into simulating post-surgical hemodynamics, assessing joint atrio-ventricular function, and quantifying vessel wall stiffness. Technical Impact - We believe that the anatomical variability of CHD is at or beyond the limits of what has been previously attempted in automated medical image analysis. Strong anatomical priors cannot be enforced, and relating information across subjects with dramatically different heart configurations is very difficult. In addition, segmentation in cardiac MR is clinically advantageous but more technically challenging than analysis of CT images (MM-WHS challenge results, Zhuang et al., 2019). Finally, although the challenge dataset is reasonably large, the set of possible defect combinations is huge, and specific configurations are often represented by a single sample. To be useful, any approach must be able to learn from images that represent only a subset of the possible anatomies, and generalize well to previously unseen configurations. Hence, we believe that the HVSMR-II challenge will be motivating to research for many years, and will encourage development of algorithms that bring automated analysis to a variety of new clinical applications involving extreme anatomical variability