TL;DR
This paper presents a deep learning approach using Fully Convolutional Networks (FCNs) to automatically segment tricuspid valve (TV) leaflets from transthoracic 3D echocardiograms (3DE) in children with Hypoplastic Left Heart Syndrome (HLHS). By incorporating additional inputs like annular curves and commissural landmarks, they achieved segmentation accuracy comparable to human experts, paving the way for faster and more precise clinical modeling and surgical planning.
What Problem Does This Paper Address?
Hypoplastic Left Heart Syndrome (HLHS) is a severe congenital heart defect where the right ventricle is the sole pump for systemic circulation, making the tricuspid valve (TV) critical. TV failure in these patients leads to heart failure, and current repair outcomes are poor. While 3D echocardiography (3DE) offers detailed valve visualization, precise modeling and quantification require accurate segmentation of the TV leaflets. However, existing methods for rapid TV segmentation are lacking, hindering clinical application for this complex pediatric population. Manual segmentation is time-consuming and prone to significant inter- and intra-user variability.
Methods
The authors employed a Fully Convolutional Network (FCN), specifically a modified V-Net architecture, for segmenting the TV leaflets. They explored various input configurations to optimize performance:
- Input Frames: They tested using only the mid-systolic (MS) frame, or including mid-diastolic (MD), end-systolic (ES), and end-diastolic (ED) frames, as well as consecutive systolic phases (CSP).
- Annular Curve Input: A representation of the TV annular curve was added as a signed distance map (SDM) channel.
- Commissural Landmarks Input: Coordinates of anterior-septal, posterior-septal, and anterior-posterior commissures were converted to SDMs and used as input channels.
- Resampling: They investigated the impact of enforcing a maximum voxel spacing and minimum leaflet segmentation height.
- Data Augmentation: Techniques like random affine transformations, contrast adjustment, and histogram clipping were used.
- Loss Function: A weighted sum of Soft Dice loss and a spatially weighted cross-entropy loss was employed to better handle thin structures and penalize boundary errors.
Datasets
The study utilized a dataset of 161 transthoracic 3DE images from 129 HLHS patients.
- Training Data: 133 image-segmentation pairs were used for training the FCN.
- Testing Data: 28 images were reserved for independent validation. The data was acquired retrospectively from an institutional database and prospectively from an IRB-approved research protocol.
Key Insights and Ideas
- Deep Learning for Pediatric TV Segmentation is Feasible: The study demonstrates that FCNs can effectively segment TV leaflets from transthoracic 3DE in a challenging pediatric population.
- Anatomical Priors are Crucial: Incorporating anatomical information like the annular curve and commissural landmarks significantly improved segmentation accuracy, bringing it closer to human inter-user variability.
- Resampling Matters: Enforcing a consistent voxel spacing and minimum leaflet height through resampling generally led to better segmentation outcomes.
- Input Frame Selection is Nuanced: While more frames can be helpful, the "Consecutive Systolic Phases" (CSP) approach, which includes frames around the MS frame, generally performed best, but the benefit wasn't always linear with the number of frames.
- Signal Dropout is a Key Challenge: The authors highlight that signal dropout in 3DE images can lead to segmentation errors (holes in leaflets), and while their method improved, it's not entirely immune.
Key Results
- Merged Segmentation: The FCN with an annular curve input achieved a median Dice Similarity Coefficient (DSC) of 0.86 [IQR: 0.81–0.88] and a Mean Boundary Distance (MBD) of 0.35 [0.23–0.4] mm.
- Individual Leaflet Segmentation: The average DSC was 0.77 [0.73–0.81] and MBD was 0.6 [0.44–0.74] mm.
- Improved Accuracy with Landmarks: Adding commissural landmarks further improved individual leaflet segmentation accuracy to an MBD of 0.38 [0.3–0.46] mm.
- Comparison to Experts: The best FCN configurations (especially with annular curve and commissural landmarks) achieved segmentation accuracy comparable to or exceeding inter-user variability among human experts.
- Speed: Scripted inference took under 6 seconds on a GPU and approximately 30 seconds on a CPU per dataset.
Limitations and Future Work
- Limited Dataset Size: The availability of 3DE images for training in HLHS patients was relatively small compared to adult populations.
- Generalizability: The current approach is specific to TV segmentation in HLHS and has not been extended to mitral valves or normal hearts due to a lack of comparable datasets.
- Signal Dropout: While improved, signal dropout in 3DE images remains a challenge that can lead to segmentation errors.
- Future Work: The authors are working on deploying the model for open translational research and plan to integrate it into 3D Slicer workflows. They suggest further optimization by exploring more input frames, incorporating 2D/3D color Doppler information to better handle signal dropout, and investigating shape-aware techniques. They also note that while their DL approach is faster than atlas-based methods, atlas-based methods might be preferable when very limited training data is available.
Summary updated