Estimation of Imaging Biomarker's Progression in Post-infarct Patients Using Cross-sectional Data

Marta Nuñez Garcia, Nicolas Cedilnik, Shuman Jia, Hubert Cochet, Marco Lorenzi, Maxime Sermesant

TL;DR

This paper introduces a probabilistic disease progression modeling (DPM) approach to track the long-term maturation of cardiac scar tissue after a heart attack, using only single-time-point imaging data. The model reveals a progressive thinning of the left ventricular wall and the late appearance of intramural fat, offering a more nuanced understanding of scar evolution than current clinical metrics.

What Problem Does This Paper Address?

Following a heart attack (myocardial infarction, MI), scar tissue forms in the heart. This scar can lead to life-threatening ventricular arrhythmias, and current clinical practice often relies on simplified risk stratification methods like Left Ventricular Ejection Fraction (LVEF) or a fixed time since the infarction. However, these methods have limitations in sensitivity and specificity. There's a significant gap in understanding the dynamic, long-term process of scar remodeling and how it relates to arrhythmogenesis. This lack of detailed insight hinders personalized treatment strategies and accurate risk assessment for patients. The paper aims to bridge this gap by developing a method to model the progression of imaging biomarkers associated with scar maturation using readily available cross-sectional data.

Methods

The core of this work is the application of Probabilistic Disease Progression Modeling (DPM). Specifically, they employ a Bayesian Gaussian Process (GP) regression framework. This framework models individual time transformations to encode pathological stage and incorporates monotonicity constraints. These constraints ensure that biomarkers evolve in a biologically plausible manner (e.g., LVEF generally decreases, while scar volume might initially increase then stabilize or decrease).

The process involves:

  1. Data Acquisition: Using both Computed Tomography (CT) and Late Gadolinium Enhanced Cardiac Magnetic Resonance (LGE-CMR) imaging from post-MI patients.
  2. Biomarker Extraction:
    • LGE-CMR: Manual segmentation of the LV wall, followed by threshold-based classification into healthy tissue, border zone, and dense scar. Biomarkers derived include LVEF, dense scar volume, and border zone volume.
  3. CT: Automated LV wall segmentation using a deep learning approach (successive U-nets). Wall thickness is computed, and then categorized into four bins representing different degrees of thinning. Fat volume is also extracted.
  4. DPM Application: The extracted biomarkers from cross-sectional data are fed into the DPM. The model estimates the "scar age" (relative time since MI) for each patient based on their biomarker values and models the temporal trajectories of these biomarkers. Monotonicity constraints are applied: increasing for fat volume and wall thinning, and decreasing for LVEF, dense scar, and border zone volume.

Datasets

The study utilized imaging data from 49 post-MI patients who underwent both CT and LGE-CMR scans at the Centre Hospitalier Universitaire (CHU) de Bordeaux, France. The paper does not introduce a new benchmark dataset but rather uses existing patient data for analysis.

Key Insights and Ideas

  • Cross-sectional data can reveal temporal dynamics: The most significant insight is that even with single-time-point data, DPM can infer a plausible progression of scar maturation. This is crucial because longitudinal data is often difficult and expensive to collect.
  • CT is a powerful tool for scar assessment: The study highlights CT's potential to outperform LGE-CMR in characterizing scar remodeling and predicting disease stage, particularly for tracking wall thinning and fat infiltration.
  • Scar maturation is a slow, multi-faceted process: The model suggests that LV wall thinning is a progressive process that occurs over years, and intramural fat appears relatively late in the maturation timeline.
  • Traditional biomarkers are less dynamic: LVEF, dense scar volume, and border zone volume showed minimal change over the modeled progression time, suggesting their limitations in capturing the dynamic aspects of scar evolution.
  • Potential for personalized risk stratification: By modeling individual patient trajectories, this approach could lead to more personalized risk assessments for arrhythmias and guide treatment decisions like ICD implantation or ablation therapy.

Key Results

  • The DPM predicted a progressive thinning of the LV wall over time.
  • Intramural fat volume was predicted to increase in a late stage of scar maturation.
  • LVEF, dense scar volume, and border zone volume trajectories remained almost constant, indicating limited dynamic change over the modeled period.
  • The maximum change time analysis (Table 2 and Figure 4) indicated that wall thinning biomarkers (especially severe thinning, <2mm) tend to show their most significant changes earlier in the relative progression timeline compared to fat volume or LVEF.
  • The study found that severe thinning (<2mm) was significantly greater in patients with arrhythmia, suggesting its predictive value. Other significant features included thickness between 3-2mm, 4-3mm, dense scar volume, LVEF, and fat volume.

Limitations and Future Work

  • Lack of Longitudinal Data: The primary limitation acknowledged is the use of cross-sectional data. This means the "progression time course" is a relative time and does not represent real years. The authors believe their methodology can still be useful for understanding dynamics and relative timing, but direct validation with longitudinal data would strengthen the findings.
  • Model Interpretation: While the model shows plausible evolutions, the exact interpretation of the "relative time" needs further validation.
  • Future Work: The authors suggest that this DPM approach can be crucial for determining when scarred tissue becomes severely thin, which is a good predictor of ventricular arrhythmia. They also propose that this method could help position patients within the scar maturation process, aiding in personalized treatment strategies like ablation therapy. Further validation with larger, longitudinal datasets is implied.

Summary updated

Abstract

International audience
BibTeX
@inproceedings{garcia2020estimation,
  author = {Marta Nuñez Garcia and Nicolas Cedilnik and Shuman Jia and Hubert Cochet and Marco Lorenzi and Maxime Sermesant},
  title = {{Estimation of Imaging Biomarker's Progression in Post-infarct Patients Using Cross-sectional Data}},
  booktitle = {M\&Ms and EMIDEC/STACOM @ MICCAI},
  year = {2020},
  doi = {10.1007/978-3-030-68107-4\_11},
}
Figure