Cardiovascular disease classification using radiomics and geometric features from cardiac CT

Technical Summary: Cardiovascular Disease Classification Using Radiomics and Geometric Features from Cardiac CT

Problem Statement
Cardiovascular diseases (CVDs) remain a leading cause of global mortality, necessitating early and accurate diagnosis. While Computed Tomography (CT) provides detailed anatomical and functional data, manual interpretation is time-consuming and prone to variability. Recent deep learning approaches for CVD classification typically operate directly on raw CT images or utilize end-to-end classifiers paired with segmentation masks. The authors argue that these methods suffer from two primary limitations: a lack of clinical interpretability and a failure to capture geometric information regarding the morphology of anatomical structures. Specifically, standard segmentation-based features do not fully represent the spatial deformations or deviations from a healthy state that may contain critical diagnostic information.

Methodology
The proposed framework decomposes the CVD classification pipeline into three distinct, interpretable stages: (i) image segmentation, (ii) image registration to a population atlas, and (iii) downstream disease classification.

1. Segmentation and Atlas Construction:
   The authors employ the Atlas-ISTN framework, which jointly learns segmentation and registration to construct a population-based anatomical atlas. To address the challenge of limited labeled data, the framework integrates pre-trained foundational models:

   • TotalSegmentator: A pre-trained model capable of segmenting 117 anatomical structures, utilized here for seven major cardiac structures (atria, ventricles, myocardium, aorta, pulmonary trunk) without fine-tuning.
   • Anatomix: A general-purpose foundational model based on a 3D U-Net, pre-trained on synthetic data and fine-tuned on the MM-WHS dataset for this specific task.
   • Baseline: A 3D U-Net trained from scratch on the same data for comparison.

2. Feature Extraction:

   • Radiomic Features: Extracted from the segmentation maps of the seven cardiac structures. These include first-order statistics (mean, variance, entropy), 3D shape descriptors (volume, sphericity), and texture matrices (GLCM, GLRLM, GLSZM, NGTDM, GLDM).
   • Geometric Features: Derived from the deformation fields generated by warping individual subject images to the normative healthy atlas. These voxel-wise displacement vectors are high-dimensional; therefore, Singular Value Decomposition (SVD) is applied to reduce the feature space and extract low-dimensional, interpretable deformation features for each structure.

3. Classification:
   A Multi-Layer Perceptron (MLP) serves as the downstream classifier. To evaluate the contribution of different feature types, three separate models were trained: one using only radiomic features, one using only geometric features, and one using a combination of both. As a baseline, a 3D ResNet-50 was trained directly on raw CT images.

Key Results
Experiments were conducted using the MM-WHS dataset for segmentation training and the publicly available ASOCA dataset (20 healthy, 20 diseased cases) for disease classification, evaluated via 5-fold cross-validation.

• Segmentation Performance: Both TotalSegmentator and the fine-tuned Anatomix model achieved high Dice Similarity Coefficients (DSC) and lower boundary errors (ASD, HD) compared to the 3D U-Net trained from scratch. Anatomix produced the most visually consistent segmentations, closely matching ground truth without the over-segmentation artifacts observed in TotalSegmentator (specifically regarding the descending aorta).
• Classification Performance:
  • Combined Features: The MLP utilizing both radiomic and geometric features achieved the highest accuracy of 87.50% (±10.21), with a recall of 90.00% and an F1 score of 88.01%.
  • Individual Features: Using only radiomic features resulted in 82.50% accuracy, while geometric features alone yielded 76.67%. This indicates that the two feature types provide complementary diagnostic information.
  • Baseline: The ResNet-50 trained directly on raw CT images achieved an accuracy of 67.50% and a recall of 60.00%, significantly underperforming the feature-based approach.

Significance and Claims
The paper claims that breaking down the classification pipeline into segmentation, registration, and feature extraction enhances clinical interpretability compared to end-to-end deep learning models. The results support the hypothesis that geometric features related to anatomical morphology (specifically deformation fields relative to a healthy atlas) contain diagnostic information complementary to standard radiomic features.

The authors assert that their approach achieves superior classification performance (an approximate 20% improvement over the raw image baseline) without relying on large-scale image datasets or end-to-end training. By leveraging pre-trained foundational models, the method demonstrates efficacy in low-data regimes. The work concludes that interpretable features derived from anatomical structures and deformation fields are more discriminative for CVD classification than features learned directly from raw intensity data. Future work is proposed to validate these findings on larger datasets and to incorporate uncertainty estimates into the pipeline to further improve clinical utility.