Abstract
Accurate classification of Atrial Fibrillation (AF) sub-types—Paroxysmal (PaAF) and Persistent (PeAF)—is critical for determining clinical intervention strategies. However, manual classification from 3D Computed Tomography (CT) volumes is challenging, and automated methods often struggle with either the lack of local spatial context (traditional radiomics) or overfitting on small medical datasets (deep learning). This paper presents RIDL (Radiomics-Informed Deep Learning), a hybrid framework that integrates local radiomic features into a deep neural network. By utilizing a novel de-correlation loss and a feature bank mechanism, RIDL ensures that the deep learning component learns information complementary to hand-crafted features. Experimental results on a dataset of 172 patients demonstrate that RIDL achieves a state-of-the-art AUC of 86.9%, significantly outperforming standalone radiomic and deep learning approaches.
1. Introduction
Atrial Fibrillation (AF) is a prevalent cardiac arrhythmia associated with increased risks of stroke and heart failure. Clinically, AF is categorized into Paroxysmal AF (PaAF), which terminates spontaneously within seven days, and Persistent AF (PeAF), which lasts longer and typically requires more aggressive treatments like catheter ablation. Distinguishing between these sub-types is essential for personalized treatment planning.
While 3D CT imaging of the left atrium provides detailed structural information, extracting predictive markers for AF sub-types remains difficult. Current automated approaches generally fall into two categories:
- Radiomics-based methods, which extract hand-crafted features (e.g., texture, shape). These are interpretable but often rely on global statistics that fail to capture subtle local tissue heterogeneity.
- Deep Learning (DL) methods, which can learn complex hierarchical features but are prone to overfitting when trained on high-dimensional 3D volumes with limited sample sizes.
To address these limitations, we propose the Radiomics-Informed Deep Learning (RIDL) framework, which treats radiomics as an "information prior" to guide and regularize the deep learning process.
2. Methodology
2.1 Local Radiomic Fusion
Unlike traditional radiomics that calculate a single feature vector for an entire volume of interest, RIDL employs Local Radiomic Fusion. The framework calculates radiomic features (such as Gray Level Co-occurrence Matrix textures) for small local patches around every voxel. This generates a spatial map of textures that is fused with low-level features from a 3D U-Net (m3DUNet). By providing the network with these "pre-digested" local texture maps, the model is less burdened with learning basic structural patterns from scratch, thereby reducing the risk of overfitting.
2.2 Feature De-correlation Loss
A common pitfall in hybrid models is information redundancy, where the deep learning component merely replicates the patterns already captured by the radiomic features. To enforce the learning of novel, complementary information, RIDL introduces a Feature De-correlation Loss. This loss function penalizes the correlation between the deep-learned features and the radiomic features, forcing the neural network to focus on signals that hand-crafted features might miss.
2.3 Feature Bank and Regularization
To accurately estimate feature correlations without requiring prohibitively large batch sizes, the framework utilizes a Feature Bank. This first-in, first-out queue stores feature statistics across multiple training iterations, providing a stable and memory-efficient way to calculate the de-correlation loss. Additionally, the backbone m3DUNet performs a self-reconstruction task, which acts as an auxiliary regularizer to improve the robustness of the learned representations.
3. Experiments and Results
3.1 Dataset and Preprocessing
The study involved 172 patients (94 PaAF and 78 PeAF) from Sun Yat-Sen Memorial Hospital. The inputs consisted of 3D CT volumes centered on the left atrium and binary masks of Epicardial Adipose Tissue (EAT). All volumes were normalized and resized to a resolution of 96x128x128. The model was evaluated using five-fold cross-validation.
3.2 Performance Comparison
RIDL was compared against traditional radiomics and standard deep learning architectures. The results indicate a clear advantage for the hybrid approach:
- Radiomics (LASSO): Achieved an AUC of 83.4%. While effective, the reliance on global features limited its ability to capture localized tissue changes.
- Deep Learning (3D U-Net): Achieved an AUC of 77.2%. The performance was hampered by overfitting on the relatively small 3D dataset.
- RIDL (Proposed): Achieved the highest performance with an AUC of 86.9% and an accuracy of 76.9%.
Ablation studies further validated the framework's components. The inclusion of local radiomic features improved the AUC by 1.6% over a standard DNN, while the addition of the de-correlation loss provided the final performance boost to reach the state-of-the-art result.
4. Discussion and Conclusion
The success of RIDL highlights the value of integrating domain-specific knowledge (radiomics) into deep learning architectures. By shifting from global to local radiomics, the model successfully captures the heterogeneity of the heart surface, which is a key indicator of AF severity. Furthermore, the de-correlation loss ensures that the hybrid model is truly greater than the sum of its parts by encouraging the discovery of non-redundant features.
Future work will focus on validating the RIDL framework on larger, multi-center datasets to ensure generalizability across different imaging protocols. Additionally, exploring automated feature selection for the local radiomic maps could further optimize the framework for various medical imaging tasks beyond Atrial Fibrillation classification.
Summary updated