Deep learning reduced order models of vaginal tear propagation

Area of Science:

• Biomedical Engineering
• Computational Mechanics
• Obstetrics

Background:

• Childbirth can cause significant maternal trauma, including vaginal tears and pelvic floor disorders.
• Current clinical methods struggle to accurately predict complications associated with vaginal delivery.
• Understanding vaginal biomechanics is crucial for preventing long-term maternal health issues.

Purpose of the Study:

• To develop novel computational methods for predicting vaginal deformations and tearing during childbirth.
• To integrate finite element (FE) analysis, proper orthogonal decomposition (POD), and machine learning (ML) for enhanced predictive accuracy.
• To create efficient computational tools for non-invasive quantification of vaginal tissue damage.

Main Methods:

• FE models of vaginal canals were created using ex vivo rodent data, simulating increasing pressure and tear propagation.
• Full-order ML models and POD-based reduced order models were developed using FE displacement field snapshots.
• ML algorithms were employed to compute coefficients for both model types, analyzing different collagen fiber organizations.

Main Results:

• Both full-order ML and POD-ML models accurately approximated FE results, with root mean squared errors of O(10^-2).
• POD-ML models demonstrated superior training efficiency (O(10) seconds) compared to ML models (O(10^2) seconds).
• Prediction times for both models were rapid (O(10^-3) seconds), enabling quick online assessments.

Conclusions:

• The integration of FE analysis, POD, and ML provides a powerful computational framework for predicting vaginal delivery outcomes.
• POD-based reduced order models offer a significant advantage in training efficiency without compromising prediction accuracy.
• These computational tools represent a promising non-invasive approach for quantifying vaginal tissue deformations and tears, aiding in the prediction and prevention of childbirth-related injuries.