From Motion to Meaning: Biomechanics-Informed Neural Network for Explainable Cardiovascular Disease Identification Journal: Computers in Biology and Medicine
Valentin Comte, Mario Ceresa, Gemma Piella, Miguel A. González Ballester · 2025
Read From Motion to Meaning: Biomechanics-Informed Neural Network for Explainable Cardiovascular Disease Identification Journal: Computers in Biology and Medicine by Valentin Comte, Mario Ceresa, Gemma Piella, Miguel A. González Ballester (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-05783.txt.
Introduction
Valentin Comte, Gemma Piella, Mario Ceresa, and Miguel González Ballester teach a neural network the physics of heart muscle to spot heart disease. Their 2025 paper registers moving heart images with deep learning, then regularizes the motion with a material model so the deformation stays physically sensible. From the fitted motion they extract local strain, how much each patch stretches, as a rich feature set.
On a public cardiac challenge data set, the alignment scored high overlap for both ventricles and the muscle wall. Five standard classifiers then used the strain features to label disease types. Because the features come from an explicit biomechanical model, doctors can inspect why a case was flagged, a property black-box systems lack.
Catalog record
| CREATOR | Valentin Comte, Mario Ceresa, Gemma Piella, Miguel A. González Ballester |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 41 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-05783
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
From Motion to Meaning: Biomechanics-Informed Neural Network for Explainable Cardiovascular Disease Identification Journal: Computers in Biology and Medicine Valentin Comte, Mario Ceresa, Gemma Piella, Miguel A. González Ballester Abstract Cardiac diseases are among the leading causes of morbidity and mortality worldwide, which requires accurate and timely diagnostic strategies. In this study, we introduce an innovative approach that combines deep learning image registration with physics-informed regularization to predict the biomechanical properties of moving cardiac tissues and extract features for disease classification. We utilize the energy strain formulation of Neo-Hookean material to model cardiac tissue deformations, optimizing the deformation field while ensuring its physical and biomechanical coherence. This explainable approach not only improves image registration accuracy, but also provides insights into the underlying biomechanical processes of the cardiac tissues. Evaluation on the Automated Cardiac Diagnosis Challenge (ACDC) dataset achieved Dice scores of 0.945 for the left ventricular cavity, 0.908 for the right ventricular cavity, and 0.905 for the myocardium. Subsequently, we estimate the local strains within the moving heart and extract a detailed set of features used for cardiovascular disease classification. We evaluated five classification algorithms — Logistic Regression, Multi-Layer Perceptron, Support Vector Classifier, Random Forest, and Nearest Neighbour — and identified the most relevant features using a feature selection algorithm. The best performing classifier obtained a classification accuracy of 98% in the training set and 100% in the test set of the ACDC dataset. By integrating explainable artificial intelligence, this method empowers clinicians with a transparent understanding of the model’s predictions based on cardiac mechanics, while also significantly improving the accuracy and reliability of cardiac disease diagnosis, paving the way for more personalized and effective patient care. 1 Introduction Cardiovascular diseases (CVDs) remain the leading cause of death worldwide, accounting for nearly 17.9 million deaths each year, which represents approximately 31% of all global deaths ( Nawaz et al. 2021 ) . The early and accurate diagnosis of CVDs is crucial for effective treatment and management, potentially reducing the mortality rate and improving patient outcomes. However, the traditional methods of diagnosing CVDs, which often rely on visual interpretation of medical images and clinical data, are time-consuming, subject to inter-observer variability, and can be prone to errors. In this context, the advent of deep learning (DL) technologies has opened new avenues for the automatic classification of CVDs. DL algorithms, particularly convolutional neural networks (CNNs), have demonstrated remarkable performance in medical image analysis, often surpassing human experts in accuracy. Cardiac disease classification is a rapidly evolving field, with advances in machine learning and DL techniques significantly improving the accuracy and efficiency of disease detection and classification. The integration of DL for the classification of CVDs offers several significant advantages. Firstly, it enables the rapid and consistent analysis of vast amounts of data, facilitating early detection and intervention. Secondly, it reduces the workload on healthcare professionals, allowing them to focus on patient care rather than repetitive diagnostic tasks. Additionally, automated classification systems can be continuously updated and improved as new data becomes available. This ensures that diagnostic accuracy is maintained over time and generalizes well across images acquired with different scanners, which may exhibit heterogeneous appearances. In that context, cine-cardiac magnetic resonance imaging (MRI) constitutes a valuable asset because of its ability to capture subtle motions of the cardiac tissues. Several cine-cardiac MRI datasets are available for research purposes, each offering different advantages and limitations. The ACDC dataset ( Bernard et al. 2018 ) is one of the most widely used benchmarks. It consists of 2D short-axis cine-MRI slices stacked to reconstruct the left ventricle volume, with expert annotations and pathology labels for 150 patients. It supports both segmentation and classification of five cardiac conditions. The UK Biobank (UKBB) dataset ( Petersen et al. 2016 ) follows a similar 2D acquisition scheme but includes a much larger and more diverse population, making it ideal for large-scale studies. In contrast, the dataset used in Galazis et al. 2023 consists of true 3D cine-MRI volumes of the left atrium acquired over time, allowing for anatomically detailed biomechanical strain analysis. However, this dataset is not publicly available and includes only a small number of subjects collected at a single centre. Building on the ACDC dataset, several studies have explored automatic CVD classification using machine learning. Isensee et al. 2018 extracted features from segmentations and trained ensembles of MLPs and random forests, achieving classification accuracies of 94% (training) and 92% (test). Cetin et al. 2018 used radiomics and SVMs with feature selection, reaching 96% and 92% accuracy, while Khened et al. 2017 proposed a DenseNet-based segmentation pipeline coupled with random forest classification, achieving 90% and 96%. Although effective, these approaches rely primarily on static anatomical features and do not leverage dynamic or biomechanical information from cine-MRI sequences. This gap can be effectively bridged using deformable image registration (DIR), which computes a nonlinear deformation field (DF) to align pairs of images. DIR is particularly advantageous for the dynamic analysis of moving tissues, offering precise approximations of particle DFs. By incorporating biomechanical priors to constrain the DF, more realistic and physiologically plausible deformations can be achieved. This approach is especially useful in analyzing the dynamic motion of the heart using cine-cardiac MRI sequences, enabling the detection of biomechanical anomalies. Such analysis plays a crucial role in diagnosing pathologies that disrupt normal cardiac tissue motion, with the added benefit of enhancing the explainability of artificial intelligence (AI) models in clinical settings. Biomechanical analysis of cardiac tissues involves examining the mechanical properties and behavior of the heart under various physiological conditions. This includes understanding the stiffness, elasticity, and motion of cardiac tissues, providing valuable insights into heart health and functionality. Since biomechanical properties can be significantly altered in various cardiac pathologies, this analysis is essential for diagnosing and treating such conditions. Advanced techniques, such as biomechanics-informed neural networks and physics-informed neural networks, have been developed to track myocardial motion and register cine-MRI scans accurately. By integrating biomechanical insights, these approaches make the diagnostic process more transparent and reliable, allowing clinicians to understand and trust the AI decision-making process. This transparency is crucial for the adoption of AI in medical practice, ensuring that the models provide clear, interpretable results that can be effectively used in patient care. Qin et al. 2020 introduced an innovative method to monitor myocardial movements by embedding a biomechanics-informed prior into their registration model, rather than using an explicit generic regularization. This approach learns a manifold for biomechanically plausible deformations. Their method was evaluated on the UK biobank (UKBB) ( Petersen et al. 2016 ) and ACDC datasets. By registering the 2D slices of the ES frames onto the ED frames, they achieved Dice scores of 0.684 for the apical slices, 0.796 for the mid-ventricle slices, and 0.725 for the basal slices. While their method outperformed traditional registration methods with conventional regularization, it was not compared to SOTA DL registration methods. In a similar inspiration, Wu et al. 2024 utilised a registration framework that integrates neural differential equations to model DFs, learning dynamics directly from data without relying on physical priors. By registering the ES onto the ED frames of the ACDC dataset, they reported Dice scores of 0.702, 0.788, and 0.793 for the apical, mid-ventricle, and basal slices,