HRA-TPF: a Hierarchical Radiomics Feature Extraction and Attention-Guided Temporal Modeling Framework for longitudinal coronary CT angiography plaque progression prediction

Longitudinal coronary computed tomography angiography (CTA) is widely used to monitor atherosclerotic plaque progression, and the primary challenge lies in extracting multi-scale discriminative features from high-dimensional imaging data and effectively modeling the temporal change pattern between two time points. This paper proposes a Hierarchical Radiomics Feature Extraction and Attention-Guided Temporal Modeling Framework (HRA-TPF) for longitudinal coronary CTA plaque progression prediction, comprising three functional modules. The Hierarchical Radiomics Feature Extraction module (HRFE) extracts imaging features in parallel at three semantic levels, namely voxel, texture, and structure, and through a two-stage joint selection combining least absolute shrinkage and selection operator (LASSO) regression and maximum relevance minimum redundancy (mRMR), compresses the 1218-dimensional raw feature pool to an 87-dimensional discriminative feature vector. The Attention-Guided Feature Refinement module (AFGR) adaptively weights the feature vector through a dual-branch channel-spatial attention mechanism, enhancing the representational capacity of discriminative features. The Temporal Modeling module (TM), composed of a two-layer Long Short-Term Memory (LSTM) network, models the refined feature pairs from the baseline ( T 0 ) and follow-up ( T 1 ) time points as an input sequence and encodes the direction and magnitude of longitudinal progression through gated memory mechanisms. On a longitudinal CT dataset of 520 patients from a single center, HRA-TPF achieves comprehensive performance of AUC 0.924, accuracy 88.5%, and F1 score 0.906, with point estimates exceeding those of seven comparison methods and an AUC margin of + 0.028 over the strongest baseline (ViT). The model parameter count (12.4 M) and single-sample inference latency (18.7 ms) are only 14.3% and 21.4% of those of ViT, respectively. Five-fold cross-validation gives a mean AUC of 0.922 ± 0.003, supporting the internal stability of the model within this cohort, though the dataset is drawn from a single institution and external validation on independent, multi-center data has not yet been performed. Experimental results indicate that HRA-TPF achieves a favorable balance between predictive performance and computational efficiency, offering a lightweight and interpretable approach to computer-aided plaque progression prediction from longitudinal coronary CT imaging.