HeartUnloadNet: A cycle-consistent graph network with reduced supervision for predicting unloaded cardiac geometry...

Area of Science:

• Computational mechanics
• Biomedical engineering
• Artificial intelligence in medicine

Background:

• The unloaded cardiac geometry is crucial for personalized biomechanical modeling but cannot be directly measured in vivo.
• Clinical imaging captures pressure-loaded states, necessitating complex reconstruction methods.
• Traditional inverse finite element solvers are computationally intensive and prone to convergence issues.

Purpose of the Study:

• To develop an efficient and accurate deep learning framework, HeartUnloadNet, for predicting unloaded left ventricular geometry.
• To enable direct prediction from clinical end-diastolic states, bypassing traditional iterative methods.

Main Methods:

• HeartUnloadNet utilizes a graph attention-based neural network incorporating mesh topology and physiological parameters.
• A cycle-consistent bidirectional training strategy allows for reduced supervision.
• The model was trained and validated on over 10,000 finite element simulations.

Main Results:

• HeartUnloadNet achieved sub-millimeter accuracy (Dice: 0.986, Hausdorff: 0.083 cm).
• The framework is over 100,000 times faster than conventional solvers, with inference times of 0.02 seconds.
• High accuracy was maintained with minimal labeled data due to cycle consistency.

Conclusions:

• HeartUnloadNet offers a scalable and accurate alternative to traditional methods for estimating unloaded cardiac geometry.
• The framework achieves real-time performance and biomechanical fidelity through mesh-aware learning and reduced supervision.
• This work supports the integration of AI into clinical workflows for patient-specific cardiac modeling.