1. Introduction

Congenital heart disease (CHD) is generally defined as a developmental defect in the structure of the heart or great vessels or both that is present in a newborn. It is the most common type of birth defect accounting for approximately one-third of all diagnosed congenital disorders [1,2]. Global birth prevalence of CHD was estimated to be 17-18 cases per 1,000 newborns annually from 1990 to 2017 [1,2]. This corresponds to 2.5 million births of neonates with CHD every year. A report in 2017 estimated that nearly 12 million people were living with CHD in the world [1]. In the United States, the birth prevalence of CHD was estimated to be 1.23% and 466,566 people were living with congenital cardiovascular defects in 2017 [3]. The most prevalent CHDs are valvuloseptal defects with abnormally-formed heart valves or septa [3,4], contributing about 50% to the total burden of CHDs.

A key process in both valvular and septal development is the proper formation of endocardial cushions in the early, tubular embryonic heart [5–7]. Endocardial cushions function as early valves in embryonic hearts, and are precursors to many critical structures in the mature heart, including the atrial and ventricular septa, and all four sets of cardiac valves [8,9]. Endocardial cushion formation begins for the chicken, a common model for studying embryonic development, at about stage 14 (Hamburger-Hamilton - HH [10]). This corresponds to the mouse at embryonic day 9, and the human at embryonic day 20 (Carnegie stage 9) [9]. At this stage, the heart is a looped tube with an outer myocardial cell layer and an inner endocardial cell layer separated by an extracellular matrix layer called cardiac jelly. Localized expansion of cardiac jelly occurs at the atrioventricular canal (AVC) and in the outflow tract (OFT) due to a poorly understood but regulated secretion of extracellular matrix by the myocardium [9]. As a result, the squamous epithelial lining of the endocardium at these two locations is displaced away from the myocardium, creating distinct endocardial cushions.

After cushion formation, mesenchymal cells are observed in AVC cushions at about HH stage 17 [11,12]. These mesenchymal cells are generated through epithelial-to-mesenchymal transition (EMT) [13]. This localized process is initiated by myocardial-derived inductive signals that are received by a subset of endocardial cells. The stimulated endocardial cells undergo cell-cell separation and transform into mesenchymal cells which then migrate into the cardiac jelly and proliferate in the cushions [14]. EMT is initiated later in the OFT cushions than in the AVC cushions. The OFT cushions begin EMT at HH stage 18 and are populated by numerous mesenchymal cells by HH stage 20. Besides endocardial-derived mesenchymal cells, cardiac neural crest-derived cells are also observed in OFT cushions. Neural crest cells migrate into the distal OFT (truncus) beginning at HH stage 18/19 [9].

The cell-populated AVC and OFT cushions subsequently undergo growth and remodeling to form valve leaflets and the septal structures of the mature heart [15–17]. The AVC cushions contribute to form the mitral and tricuspid valves, as well as the ventricular septum and the atrioventricular septum. The mesenchymal cells in OFT cushions are thought to contribute the aortic and pulmonary valves. The neural crest cells in OFT cushions contribute to form the aorticopulmonary septum.

Cushion formation is a complex developmental process, mediated by many molecular and cellular activities [9]. It is likely that subtle perturbations in early endocardial cushion development can result in defects in cardiac valves and septa, the most common types of CHDs. Abnormal heart valves and some cardiac septal defects can be traced to the malformation of endocardial cushions during heart development [7,18].

Many studies that investigated the genetics and molecular mechanisms of cushion formation [19–28] and the effects of environmental perturbations such as ethanol exposure [29–33] and alterations in blood flow dynamics [34–37] on cushion development, involve measuring cushion size and the cushion cellularized fraction to quantify the progression of cushion development. Most of these studies imaged serial sections of embryonic hearts using histology or immunohistochemistry methods [19,20,34,35], then manually counted the number of cells to quantify the cell density of cushions. Optical coherence tomography (OCT) is another technique for imaging the developing heart used in many studies. OCT can acquire micro-level spatial resolution images with a 1–2mm imaging depth in embryonic tissues. Compared with conventional histology and immunohistochemistry methods, OCT can image larger 3D samples at a much higher speed.

We have previously used OCT to study the effects of prenatal alcohol exposure on heart development and found that ethanol-exposed embryos had smaller endocardial cushions than controls [30] and that co-administration of betaine [31], folic acid [32], or glutathione [33] prevented this reduction. In these studies, endocardial cushions were measured by manually segmenting 3D OCT images. A recent study [38] tracked EMT progression in OFT cushions by counting mesenchymal cells in a 2D OCT image. In that study, the investigators manually delineated cushion boundaries from one image plane of the OFT and then counted cell regions (separate clusters of bright pixels) in the cushions using the Blob Analysis function provided by MATLAB (MathWorks, MA, U.S.). Manual segmentation of endocardial cushions from 3D OCT images is time-consuming and labor-intensive. In our experience, it usually takes an expert with much practice around 40 minutes to label one OCT heart volume by hand and a person with less experience will spend even longer time. The manual analysis limits the throughput of experiments especially when large cohorts of embryos are required. Therefore, a method that can automatically measure the cushion size and EMT progression is needed.

The aim of the current study was to develop and demonstrate a method to automatically measure the size of endocardial cushions and the fraction of the cushions occupied by cells from 3D OCT images. We used a two-stage deep learning framework [39] to first detect the location of the heart, and then segment the endocardial cushions. The acellular cardiac jelly and the cell-populated cushion regions were detected by K-means clustering [40,41]. The proposed method can provide the 3D cushion structure, the cushion volume, and the cushion cellularized fraction, automatically in less than one minute. Application of this rapid 3-D method produced novel insights regarding the looping heart.

2. Materials and methods

2.1 Sample preparation

Fertilized quail eggs (Coturnix coturnix communis; Northwest Heritage Quail, Pullman, WA) were incubated in a humidified incubator (G.Q.F. Manufacturing, Savannah, GA) at 38°C. After 50-72 hours of incubation (HH stage 15-20), embryos were dissected, and fixed in 10% formalin solution (Sigma-Aldrich, MA, U.S.). Fixed embryos were then washed and stored in PBS solution (Sigma-Aldrich) at 4°C prior to imaging.

2.2 Image acquisition

The hearts were imaged using a custom-built, spectral-domain OCT system used previously for similar studies [32,33,36,42]. The system has a light source centered at 1310 nm and a 75 nm full-width at half-maximum bandwidth. The line rate of the OCT system is 47 kHz. For each heart, a 3D volume (1000 lines/frame, 1000 frames/volume, total volume 2.0 (L) ×2.0 (W) × 3.4(H) mm) was recorded. During imaging, the embryos of HH stages 15-18 were placed in PBS solution, and the larger embryos at HH stage 19-20 were cleared in formamide solution (Sigma-Aldrich, MA, U.S.) for reduced scattering and deeper penetration. We imaged 15 embryos per HH stage and a total of 90 image volumes were captured.

2.3 Image processing

2.3.1 Overview of image processing

The endocardial cushions initially contain no cells, only cardiac jelly. Later, mesenchymal cells appear in the cardiac jelly adjacent to the lumenal endocardium through EMT. These mesenchymal cells then proliferate and migrate away from the endocardium deeper into the cardiac jelly. Besides mesenchymal cells, cardiac neural crest cells are also observed in OFT cushions. Eventually, the cushions are largely populated by cells. The goal of this study was to develop and demonstrate methods to characterize the morphology of endocardial cushions by measuring the total cushion volume V_cushion and the cushion cellularized fraction F_cell by segmenting the acellular and cellular cushion regions from OCT images. F_cell is defined as V_cell / V_cushion, where V_cell is the volume of cellular cushion regions. V_CJ is the volume of cardiac jelly. In OCT images (e.g. Figure 1), the acellular cardiac jelly has a low intensity signal close to background levels because cardiac jelly is transparent and scatters very little light. On the other hand, cushion regions populated by cells have relatively higher intensity signals close to that of other embryonic tissues. In this study, the endocardium was segmented together with endocardial cushions because it cannot be distinguished from cushion cells in OCT images (Fig. 1(b)). However, this will not affect our analysis because the endocardium is only a monolayer of epithelial cells at the stages investigated in this study.

 

Fig. 1. An exemplar cross-sectional OCT image of an embryonic heart at HH stage 17. (a) The image annotated with different cardiac segments. AVC: atrioventricular canal; OFT: outflow tract. (b) The image annotated with cardiac tissues: myocardium (myo), endocardium (endo) and endocardial cushion components including cardiac jelly (CJ) and mesenchymal cells (MCs). The regions enclosed by the red contours will be segmented, including MCs, acellular CJ and endocardium.

Download Full Size | PDF

The overall pipeline of processing OCT images is shown in Fig. 2. The first step is detecting the heart within the original image volume (Fig. 2(a)) to crop a 3D region of interest (ROI) (Fig. 2(b)). The original image volumes include considerable non-cardiac voxels (Fig.2a). Defining an ROI is useful to reduce memory usage and computation time. In the second step, endocardial cushions are extracted from the cropped ROI volume (Fig. 2(c)). At this step, total cushion volume V_cushion can be calculated (Fig. 2(d)). Finally, the cushion voxels are partitioned into two groups: the low-intensity acellular cardiac jelly (blue regions in Fig. 2(e)-(f)) and the high-intensity cell-populated regions (yellow regions in Fig. 2(e)-(f)), from which the cushion cellularized fraction can be computed. All of these steps were performed in 3D. The procedures for each step are described in greater detail below.

 

Fig. 2. The flowchart of cushion segmentation and cellularization quantification from an OCT image volume. (a) the original OCT image volume (log-transformed). (b) 3D ROI enclosing the heart was automatically detected as indicated by the red rectangle. (c) 3D segmentation of cushions from the cropped ROI volume. The red contours are boundaries of segmented cushions. (d) 3D rendering of the segmented cushions. (e) Cushion voxels were clustered into two groups: acellular cardiac jelly (blue) and cell-populated regions (yellow). (f) 3D rendering of clustered cushion voxels showing spatial patterns of the acellular (blue) and cellular (yellow) cushion regions. ROI: region of interest.

Download Full Size | PDF

2.3.2 ROI detection and cushion segmentation via deep learning

We achieved ROI detection and cushion segmentation using a two-stage V-Net framework (Fig. 3). V-Net is a fully convolutional network architecture designed for volumetric image segmentation [39].

 

Fig. 3. The two-stage neural network framework proposed for ROI detection and cushion segmentation. ROI-detection V-Net and cushion-segmentation V-Net had the same architecture. ROI-detection V-Net was trained with lower resolution images for localizing the heart. Cushion-segmentation V-Net was trained with higher resolution images for fine cushion segmentation.

Download Full Size | PDF

In the ROI detection step, the original image volume (1000*512*1000) was first down-sampled to a fixed size (80*128∗80, resolution: 25*24*25 µm), then log-transformed and normalized. The log-transformation was applied here because our original OCT datasets were stored in a linear scale format. The preprocessed image volume was fed into the ROI-detection V-Net to obtain a rough cushion mask. Subsequently, we up-sampled the predicted mask back to the original size (1000*512*1000). The ROI (red box in Fig. 1) with edges 150 µm away from the binary cushion mask was then cropped from the original image.

In the cushion segmentation step, the cropped ROI volume was first resized, log-transformed and normalized. Then the preprocessed image volume (fixed size: 128*128*128) was fed into a trained model: the cushion-segmentation V-Net to predict labels of cushions. The architecture of the two V-Nets was the same and adapted from Milletari et al. [39]. Generalized Dice overlap [43] was used as the loss function to mitigate the issue of label imbalance.

In our experiments, we separated image volumes of 90 embryonic hearts into three groups: embryos at HH stage 15-16, HH stage 17-18, and HH stage 19-20, respectively. Distinct models were trained for each group to accomplish the two steps described above. For each group, 24 samples were randomly selected for training and 6 for testing. The 6 test samples had never been seen in the training, to provide an unbiased evaluation of the performance of the final trained models. During training, the 24 training samples were used for 5-fold cross-validation [44], to reduce the potential of over-fitting caused by the small sample size. Specifically, the 24 training samples were divided into five subgroups, to train five models. At each training round, one subgroup was selected as the validation set and the remaining four subgroups were used for training. The final cushion mask is generated by binarizing the average of the probability maps output from the five models with a threshold 0.5. Moreover, data augmentation including random reflection, translation, rotation, and scaling was adopted during training, specifically, the range of random rotation was [−15, 15] degrees, the range of translation was [-10, 10] voxels, and the range of scaling was [0.8, 1.2].

We adopted Adam optimizer [45] for training. The learning rate was initialized to 1e-3 and decayed by the polynomial strategy. Both V-Nets were implemented using Matlab (MathWorks, MA, U.S.) and were trained on an Nvidia Tesla P100 GPU processor for 500 epochs with a batch size of 1.

2.3.3 Cushion cellularization quantification

After binary masks of the cushions were predicted, cushion voxels were classified into two clusters: low-intensity acellular cardiac jelly and high-intensity cellular regions, by a threshold. The procedure of separating cushion voxels is shown in Fig. 4. The image volume was first smoothed by a median filter with a kernel size [3*3*3] to reduce noise. Subsequently, a threshold was estimated by the K-means algorithm [41,46] to separate cushion voxels. We found that a bi-level threshold (cluster number = 2) performed well for hearts at HH stage 15-17, with only 1 (total 45) sample requiring manual correction. However, for older hearts at HH stage 18-20, this bi-level threshold misclassified cell voxels into the acellular cardiac jelly group. These voxels usually had reduced intensities due to the light attenuation deep in large tissue samples. Another reason for this misclassification was that EMT, cell migration and proliferation at these stages progressed rapidly, leading to uneven cell densities in the cushions. Areas of low cell density showed reduced intensities. To overcome this problem, we separated cushion voxels into more than two clusters, then set the cluster with the lowest intensity as acellular cardiac jelly and the voxels in the remaining clusters were designated as cell-containing cushion regions. The K-means cluster number we used was 3 for samples at HH stage 18 and was 4 for hearts at HH stage 19-20. This strategy worked well for all but 6 of the 45 samples which required manual refinement. We developed an interactive tool – a Matlab GUI, to visualize the classification results. This tool allowed manual adjustment of the threshold if the automated result was unsatisfactory.

 

Fig. 4. The flowchart of classifying cushion voxels. The image volume was first denoised by a median filter, then the cushion voxels were clustered into two groups by the K-means algorithm.

Download Full Size | PDF

3. Results

3.1 Endocardial cushion segmentation

Endocardial cushions in an OCT image volume were extracted by a cascade of two neural networks with the first V-Net for localizing the heart and a second V-Net for fine cushion segmentation. During training, 5-fold cross-validation was used to avoid over-fitting. Specifically, we separated the training dataset into five sub-folds, and repeated the hold-out training five times in which each sub-fold was used as the validation dataset and other four sub-folds were used for training. The 5-fold cross-validation generated five models and each model could output a prediction for new test data. The final segmentation was achieved by binarizing the average of all five predictions. Table 1 shows the segmentation results of the 5-fold cross-validation. The segmentation was evaluated by the Dice score (the overlap size of the segmentation and the manual ground truth divided by the total size of both). The higher the Dice score (range 0 to 1), the better the segmentation. Each row represents the results of an individual HH-staged group. Each value is the average Dice score of all six test hearts in an individual group. Columns from ‘Fold 1’ to ‘Fold 5’ are results of the individual five models generated by each hold-out training. The last column is the averaged output of the five models and shows highest Dice scores indicating best segmentation performance.

Table 1. Segmentation performance of 5-fold cross-validation evaluated by Dice score. The final segmentation is achieved by averaging the predictions of five folds.

View Table

A comparison of the final averaged predictions with the manual ground truth is shown in Fig. 5. The average Dice score of 18 test samples was 0.92, and the predicted labels of the sample with the median Dice score agreed well with the manually labeled ground truth (Fig. 5(b)). Even the sample with the worst Dice score (0.86, Fig. 5(c)) showed a good match between the manual ground truth and the V-Net prediction. The low Dice score of this sample was mainly caused by the inaccuracy of the manual segmentation which contained part of the myocardium (green regions in Fig. 5(c)).

 

Fig. 5. Comparison of V-Net-predicted cushion mask with manually-segmented ground truth. The box plot shows the Dice score of 18 test hearts. (a) (b) (c) show orthogonal slices of the heart volumes with the best Dice score (dot a in the box plot), the median Dice score (dot b in the box plot) and the worst Dice score (dot c in the box plot), respectively. Images are overlaid by manual mask (green) and V-Net-predicted cushion mask (red). The yellow color indicates the overlap between the manual mask and the predicted mask. The three image planes are in the order of xz, yz, xy.

Download Full Size | PDF

We compared the cushion volumes of 18 test samples computed from V-Net-predicted masks to those calculated from manual segmentation. Figure 6(a) shows a high correlation between the two measurements (Pearson's correlation coefficient = 0.978). Additionally, the V-Net results correctly predicted the difference in cushion volumes between three groups of embryos: the embryos at HH stage 19-20 had the largest cushions (red asterisks in Fig. 6(a)) and the youngest group of embryos at HH 15-16 stage had the smallest cushions (magenta circles in Fig. 6(a)). Figure 6(b) shows the Bland-Altman analysis of the two measurements. The line of equality (V-Net - Manual = 0) is within the 95% confidence interval of mean difference, illustrating no significant systematic difference between the two measurements.

 

Fig. 6. Comparison of cushion volumes estimated using V-Net segmentation versus manual segmentation. (a) the scatter plot of cushion volumes computed from the V-Net predicted mask (y-axis) versus cushion volumes measured from manually-segmented ground truth (x-axis). PCC: Pearson correlation coefficient. (b) Bland-Altman plot of cushion volumes computed from V-Net-predicted and manual masks. HH: Hamburger-Hamilton.LoA: limits of agreement. CI: 95% confidence interval.

Download Full Size | PDF

3.2 Cellularization quantification of endocardial cushions

During the development of endocardial cushions, the cardiac jelly is gradually populated by cells. To quantify the progression of cushion development, acellular cardiac jelly needs to be differentiated from cellular cushion regions. After segmenting entire cushions from OCT image volumes, the acellular cardiac jelly and the cellular cushion regions were separated using thresholds estimated by the K-means algorithm. Figure 7 shows the segmented acellular and cellular cushion regions of six representative embryonic hearts at different HH stages. The 3D renderings were created using Amira (Thermo Fisher Scientific, MA, U.S.).

 

Fig. 7. Detection of acellular and cellular cushion regions from six embryonic hearts at different stages shown in (a)-(f), respectively. The top row of each panel shows orthogonal slices of the heart overlaid by cushion boundaries (magenta lines). The middle row shows acellular (green) and cellular (red) cushion regions. The bottom row shows the 3D rendering of cellular (red) and acellular (green) cushion regions with the left showing the ventral view of the heart and the right showing the dorsal view of the heart. The cushion cellularized fraction of each heart is 34% (a), 54% (b), 63% (c), 78% (d), 87% (e) and 96% (f), respectively. HH: Hamburger-Hamilton; A: atrium; S: superior atrioventricular cushion; I: inferior atrioventricular cushion; LV: left ventricle; RV: right ventricle; C: conus; T: truncus.

Download Full Size | PDF

By observing the distribution pattern of acellular and cellular cushion regions at different stages (HH15-20), we found an asymmetric distribution of acellular cardiac jelly, which could be traced back to an early stage before EMT occurred. As shown in Fig. 7(a), at HH stage 15, the ventral side of the heart had more acellular cardiac jelly than the dorsal side. This ventral/dorsal asymmetry continued with the progression of mesenchymal cell expansion into the endocardial cushions (as shown in Fig. 7(a)-(d)), then transitioned into an asymmetry between the inner and outer curvature (as shown in Fig. 7(e)-(f)), as the ventral side bent towards the inner curvature and the dorsal side became the outer curvature, beginning at HH stage 17.

To quantify the asymmetry in acellular cardiac jelly between the heart ventral and dorsal sides, we measured the volumes of the acellular and cellular cushion regions of 18 hearts at different stages from HH15 to HH20 (three hearts per HH stage). For every heart, the ventral-side and dorsal-side cushions of the whole heart were manually separated. The delineated ventral and dorsal cushions of two exemplar hearts are shown in Fig. 8(a2), 8b2. Our measurements showed that, as the HH stage increased, in the whole heart, the ventral-side and dorsal-side cushions grew in size (Fig. 9(a)), the acellular cardiac jelly volume reduced (Fig. 9(b)), and the cellular volume (Fig. 9(c)) and the cellularized fraction (Fig. 9(d)) increased. We then focused on the acellular and cellular volumes of the AVC cushions alone. At the AVC, the superior (ventral) cushion (S in Fig. 7) is associated with the heart inner curvature and the inferior (dorsal) cushion (I in Fig. 7) is associated with the heart outer curvature. The manually delineated superior and inferior AVC cushions of two exemplar hearts are shown in Fig. 8(a3), 8b3. As expected, in the AVC cushions, with the increase in HH stage, the total cushion size increased (Fig. 9(e)), the volume of the acellular cardiac jelly reduced (Fig. 9(f)), the cellular volume (Fig. 9(g)) and the cellularized fraction (Fig. 9(h)) increased.

 

Fig. 8. 3D reconstructions of acellular/cellular cushion regions, ventral/dorsal side cushions, and the superior/inferior AVC cushion of an HH16 heart (a1-a3) and an HH19 heart (b1-b3). (a1) and (b1) show the acellular and cellular cushion regions (left: heart ventral view; right: heart dorsal view). (a2) and (b2) are surface renderings of ventral and dorsal side cushions (left: heart ventral view; right: heart dorsal view). (a3) and (b3) are surface renderings of superior and inferior AVC cushions. HH: Hamburger-Hamilton; AVC: atrioventricular canal; A: atrium; LV: left ventricle; RV: right ventricle; C: conus; T: truncus.

Download Full Size | PDF

 

Fig. 9. Measurements from the cushions in the whole heart or at AVC versus the HH stage of 18 embryos. At each HH stage, the marker shows an average measurement of three hearts and the error bar represents standard error. (a) the volumes of the ventral-side and dorsal-side cushions in the whole heart. (b) the acellular cardiac jelly volume of the ventral-side and dorsal-side cushions in the whole heart. (c) the cellular cushion volume of the ventral and dorsal cushions in the whole heart. (d) the cellularized fractions of the ventral-side and dorsal-side cushions in the whole heart. (e) the volumes of the superior and inferior AVC cushions. (f) the acellular cardiac jelly volume of the superior and inferior AVC cushions. (g) the cellular cushion volume of the superior and inferior AVC cushions. (h) the cellularized fractions of the superior and inferior AVC cushions. HH: Hamburger-Hamilton; AVC: atrioventricular canal.

Download Full Size | PDF

The measurements from individual hearts are shown in Fig. 10 and Fig. 11. In every examined heart, the ventral-side cushions had more acellular cardiac jelly than the dorsal-side cushions (Fig. 10(a) and Fig. 11(a)). At early stages of cushion development from HH15 to HH18, the cellular volumes of the dorsal-side and ventral-side cushions are similar (Fig. 10(b) and Fig. 11(b)). At late stages of cushion development (HH19-20), the cellular volumes of dorsal cushions of three hearts exceeded that of ventral cushions (Fig. 10(b) and Fig. 11(b)). In all examined hearts at different stages, the superior AVC cushion contained more acellular cardiac jelly than the inferior AVC cushion (Fig. 10(c) and Fig. 11(c)). The cellular volumes of the two AVC cushions were very close in hearts at HH stage 15-18 (Fig. 10(d) and Fig. 11(d)). In three hearts at HH stage 19-20, the cellular volume of the inferior AVC cushion was larger than that of the superior AVC cushion (Fig. 10(d) and Fig. 11(d)). In another two hearts at HH stage 19-20, by contrast, the cellular volume of the inferior AVC cushion was smaller than that of the superior AVC cushion (Fig. 10(d) and Fig. 11(d)). Figure 10(e) and Fig. 11(e) show the comparison of ventral and dorsal side OFT cushions in acellular cardiac jelly volume. The asymmetry that the ventral-side cushions have more acellular cardiac jelly than the dorsal-side was observed in the OFT regions of 15 out of 18 examined hearts. Figure 10(f) and Fig. 11(f) show the size of cellular cushion regions in ventral- and dorsal-side OFT cushions. Similar to AVC cushions, the cellular volumes of the two sides of the OFT cushions were close at early stages of cushion development (HH stage 15-18).

 

Fig. 10. (a) Scatter plot of the acellular cardiac jelly volume of dorsal cushions versus that of ventral cushions in the whole heart for 18 examined samples at different HH stages (3 per HH stage). (b) Scatter plot of the cellular volume of dorsal cushions versus that of ventral cushions in the whole heart. (c) Scatter plot of the acellular cardiac jelly volume of the inferior AVC cushion versus that of the superior AVC cushion. (d) Scatter plot of the cellular volume of the inferior AVC cushion versus that of the superior AVC cushion. (e) Scatter plot of the acellular cardiac jelly volume of the dorsal-side OFT cushion versus that of the ventral-side OFT cushion. (f) Scatter plot of the cellular volume of the dorsal-side OFT cushion versus that of the ventral-side OFT cushion. HH: Hamburger-Hamilton; AVC: atrioventricular canal.

Download Full Size | PDF

 

Fig. 11. The acellular or cellular volumes of the whole heart cushions, AVC cushions or OFT cushions of 18 examined hearts at different HH stages. (a) The volumes of the acellular cardiac jelly in the ventral and dorsal cushions. The 18 hearts were separated into three groups: HH15-16, HH17-18 and HH19-20. In a group, there are 6 samples and each dot represented an individual heart. The acellular volume of the ventral-side cushions (V) and that of the dorsal-side cushions (D) calculated from a same heart is connected by a line. *P < 0.05, **P < 0.01, ***P < 0.001. (b) The volumes of the cellular cushion regions in ventral and dorsal cushions. (c) The volumes of the acellular cardiac jelly in the superior (S) and inferior (I) AVC cushions of hearts. ***P < 0.001. (d) The volumes of the cellular cushion regions in the superior (S) and inferior (I) AVC cushions of hearts. (e) The volumes of the acellular cardiac jelly in the ventral (V) and dorsal (D) OFT cushions. ***P < 0.001. *P < 0.05 (f) The volumes of the cellular cushion regions in the ventral (V) and dorsal (D) OFT cushions.

Download Full Size | PDF

4. Discussion

Segmentation of endocardial cushions is usually the first step in characterizing early cushion development using imaging. Because manual annotation is time-consuming and becomes impractical when analyzing large numbers of samples and multiple cohorts, we developed a rapid automated segmentation of endocardial cushions from OCT images using deep learning techniques. We separated data from HH stages 15-20 into three groups of stages and for each group, a cascade of two V-Nets [39] were trained for heart localization and cushion segmentation (Fig. 3). The predicted cushion labels using this method agreed well with the manual ground truth (Fig. 5) and would be reliable for measuring cushion volume (Fig. 6).

We also investigated training a unified segmentation model on data from all of the studied HH stages. However, we found that the mean dice score of test data generated by the unified model is 0.87, lower than that of the separately-trained models (mean Dice = 0.92, shown in Fig. 5). The reason may be that the developing heart changes significantly between stages and the unified model trained with our current limited data set is not general enough to cover that large variance. Therefore, a better segmentation is achieved when training individual models specifically for each stage-group.

To quantify the progression of endocardial cushion development including the EMT process, characterizing the cellularized fraction is needed. We classified the cushion voxels into two groups: the low-intensity acellular cardiac jelly and the high-intensity cellular cushion regions, using the K-means algorithm. As shown in Fig. 7, acellular cardiac jelly and cellular cushion regions at different stages could be extracted and visualized in 3D.

By measuring the acellular and cellular cushion volumes of hearts at different stages, we found an asymmetry in the amount of cardiac jelly between ventral-side and dorsal-side cushions which had not been reported by previous studies. In all examined heart samples at different stages during cushion development, the volume of acellular cardiac jelly in ventral-side cushions was always larger than that in dorsal-side cushions (Fig. 10(a) and Fig. 11(a)). The heterogeneity of acellular cardiac jelly may signify that the environment of the ventral and dorsal sides differs in the quality or quantity of factors or extracellular matrix components that regulate EMT, cell migration or proliferation. We know that the inner curvature is subject to more shear force on the endocardial surface [47,48,49]and that the myocardium of the inner curvature has a distinct transcriptional profile that reflects the molecular signature originally found in the linear heart tube myocardium [50]. Some of these features of the inner curvature may explain the thicker acellular cardiac jelly in the endocardial cushions of the inner curvature of the looping heart.

A previous study (Moreno-Rodriguez, et al. [51]) investigated the cell distribution within the AVC cushions by counting cell nuclei from confocal micrographs of developing chicken hearts. Moreno-Rodriguez, et al. found that during cushion development, the inferior AVC cushion invariably contained more mesenchymal cells than the superior AVC cushion at different stages (HH16-20). Instead of counting the number of mesenchymal cells, we computed the size of the cellular regions of the two AVC cushions because individual cells were not resolved in OCT images. Our measurements showed that in all examined hearts at different stages, the superior cushion contained more acellular cardiac jelly than the inferior cushion (Fig. 9(f), Fig. 10(c) and Fig. 11(c)) while there was no significant difference between the volumes of cell-populated regions inferior and superior AVC cushions (Fig. 9(g), Fig. 10(d) and Fig. 11(d)). This result is not necessarily in conflict with the finding of Moreno-Rodriguez, et al. because we measured the volumes of the cell-containing regions, not the cell density. The asymmetry in the amount of acellular cardiac jelly in the AV cushions has not been reported by previous studies.

One limitation of the presented method is that cell density is not measured because individual cells are not resolved by the OCT system. The distribution pattern of cellular and acellular cushion regions is mapped in 3D, but cell density may vary spatially and over development time, and this information is not captured. In addition, the endocardial cells at the lumen are not resolved from adjacent cushion cells. To measure cushion cell density or differentiate cell types, higher resolution microscopy could complement the methods presented here. For example, automated nuclei detection has been demonstrated with 3D confocal microscopy using a fully convolutional regression network [52].

5. Conclusion

In conclusion, the methods proposed in this study achieved rapid automated segmentation of endocardial cushions and characterization of cushion components during endocardial cushion development. The output measurements of cushion volume and the cellularized fraction will facilitate future studies investigating the effects of teratogens, such as ethanol or nicotine, on cushion development. These studies usually require a large number of samples for comparison across multiple cohorts. By analyzing the 3D spatial patterns of acellular and cellular cushion regions provided by these methods, we found a spatial asymmetry of acellular cardiac jelly in endocardial cushions across all stages of cushion development, which has not been reported before. In addition, this 3D morphological pattern could be registered to other volumetric measurements that we have previously obtained, such as hemodynamics [30,53,54], shear stress [48], cardiac conduction [42,55], and molecular expression [56]. This will enable a spatial correlation of structural, functional, and molecular aspects in embryonic hearts and will facilitate a better understanding of the interplay between these factors during heart development.