Phase-Independent Latent Representation for Cardiac Shape Analysis

Josquin Harrison, Marco Lorenzi, Benoit Legghe, Xavier Iriart, Hubert Cochet, Maxime SermesantMICCAI 2021 · DOI

TL;DR

This paper introduces a novel meta-learning framework to analyze cardiac shape, specifically the left atrium (LA), by creating a phase-independent latent representation. This approach allows for the joint analysis of heterogeneous datasets (like systole and diastole cardiac images) to identify new biomarkers for predicting stroke risk in atrial fibrillation patients.

What Problem Does This Paper Address?

Atrial fibrillation (AF) is a major risk factor for stroke, and accurately stratifying this risk is crucial for patient management. While imaging data holds promise for identifying new biomarkers, analyzing it is challenging due to several factors:

  • Heterogeneity: Cardiac images are often captured at different points in the heart cycle (systole vs. diastole) and using various imaging systems, making direct comparison difficult.
  • Data Scarcity: Building complex predictive models is hindered by limited, often small, datasets.
  • Complex Anatomy: The left atrium (LA) and its appendages (LAA) are key areas for clot formation, but their shape and position vary significantly between individuals, making quantitative analysis complex.

Existing methods struggle to integrate information across these diverse datasets, limiting the development of robust, data-driven biomarkers for stroke risk prediction.

Methods

The authors propose a two-pronged approach:

  1. Graph Representation of the Left Atrium (LA): They represent the LA as a graph, focusing on the positions of the pulmonary veins (PVs) and the left atrial appendage (LAA). This simplifies the complex 3D shape into a more manageable structure for analysis. This involves:
  2. Meshing: Using the MMG4 software to create a mesh of the LA from segmentation masks.
  3. Atlas-based Registration: Creating a mean LA shape (atlas) and registering all patient shapes to this atlas to establish a consistent coordinate system.
  4. Label Warping: Warping anatomical labels (PVs, LAA) from the atlas to individual patient meshes.
  5. Graph Construction: Extracting key points (centers of mass of PVs, LAA, and junctions) as graph nodes and connecting them to form the graph.
  6. Meta-Learning Framework for Heterogeneous Datasets: They introduce a meta-learning approach that combines Multi-task Learning (MTL) and Meta Learning (ML). This framework aims to learn a common latent representation across different datasets (e.g., systole and diastole images).
  7. Variational Inference: They use variational inference to learn a shared latent space where data from different sources is represented consistently.
  8. Joint Optimization: The model is trained jointly across datasets, forcing encoders to learn a common representation that is robust to variations between datasets.
  9. Classifier: A shared decoder acts as a classifier on this latent representation to predict outcomes (e.g., thrombosis).

This meta-learning approach allows the model to leverage the strengths of multiple datasets, improving generalization and interpretability.

Datasets

The study utilizes a clinical database of 107 patients with atrial fibrillation, acquired from the Bordeaux University Hospital. This dataset is split into two subsets based on the cardiac cycle phase:

  • Systole: 50 patients (27 Thrombus positive)
  • Diastole: 57 patients (37 Thrombus positive)

The "Thrombus positive" label is a composite criterion based on CT scan detection of LAA thrombus and/or past history of embolism.

Key Insights and Ideas

  • Phase-Independent Analysis is Crucial: Treating systole and diastole images as separate datasets and jointly analyzing them significantly improves model performance and interpretability compared to analyzing them independently or as a single, unsplit dataset. This highlights that the cardiac cycle phase contains critical, distinct information.
  • Meta-Learning for Data Fusion: The meta-learning framework effectively bridges the gap between heterogeneous datasets by learning a shared latent space. This allows the model to learn from diverse data sources while maintaining a coherent representation.
  • Graph Representation Simplifies Complex Anatomy: Representing the LA as a graph of key anatomical landmarks (PVs, LAA) provides a simplified yet informative representation that captures crucial geometric variations relevant to thrombosis.
  • Interpretability Through Latent Space: The learned latent space is not just a black box; the authors demonstrate its interpretability by identifying specific anatomical features (e.g., PV ostia orientation, LAA angulation) that are important for predicting thrombosis.

Key Results

  • Improved Accuracy: The joint analysis approach achieved significantly higher test accuracies: 0.89 overall, 0.92 for diastole, and 0.86 for systole. This is a substantial improvement over independent classification (best of 0.65) or analyzing the complete dataset without splitting (0.84).
  • Robustness to Mode Collapse: Independent classification attempts suffered from mode collapse, whereas the joint analysis demonstrated robustness.
  • Enhanced Biomarker Identification: The model successfully identified key anatomical features associated with thrombosis. For instance, at diastole, PV ostia proximity and orientation were highlighted. At systole, the focus shifted to PV and LAA angulation.
  • Latent Space Coherence: Visualizations of the latent space showed clear separation of classes and a common separator for systole and diastole subsets when trained jointly, indicating a well-structured and interpretable representation.

Limitations and Future Work

  • Dataset Size and Heterogeneity: While the method shows promise, the authors acknowledge that applying it to datasets with even greater heterogeneity (different imaging modalities, patient populations) is a future direction.
  • Interpretability of Deep Encoders: The authors believe their meta-learning approach makes deep neural networks viable as encoders while retaining interpretability, suggesting further exploration of this synergy.
  • Lightweight Graph Representation: The authors suggest that this lightweight graph representation can be integrated into more comprehensive, multi-disciplinary studies of the LA.

Summary updated

BibTeX
@inproceedings{harrison2021phase,
  author = {Josquin Harrison and Marco Lorenzi and Benoit Legghe and Xavier Iriart and Hubert Cochet and Maxime Sermesant},
  title = {{Phase-Independent Latent Representation for Cardiac Shape Analysis}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2021},
  doi = {10.1007/978-3-030-87231-1\_52},
}
Figure