IC-TCA-Net: a multimodal fusion framework with dimensionality transformation for two-year recurrence prediction in NSCLC.

Abstract

Accurate prediction of recurrence in non-small cell lung cancer (NSCLC) patients after curative surgery is essential for selecting appropriate adjuvant treatments and improving long-term survival. However, current recurrence prediction models are limited by their reliance on single-modality data or fusion techniques that are insufficient for addressing the challenges of modality heterogeneity and structural imbalance. In this study, we propose an end-to-end multimodal deep learning framework that integrates preoperative CT imaging and clinicopathological variables through a dimensionality transformation module to improve two-year recurrence prediction in NSCLC patients. We retrospectively analyzed 166 NSCLC patients who underwent curative surgical resection between January 2017 and December 2020. We developed IC-TCA-Net (Integrated Clinical data with Tumor-Centric Attention Network), a novel multimodal fusion framework that combines preoperative CT images and clinicopathological variables. The model employs modality-specific subnetworks: a dual-branch ResNet-18 with tumor-centric attention modules for CT images, and a feature selection for clinicopathological variables using univariate and multivariate analyses. To address dimensional disparity between high-dimensional imaging features and low-dimensional clinical vectors, we designed a dimensionality transformation module consisting of condensation and expansion layers. Model performance was evaluated using five-fold cross-validation. IC-TCA-Net achieved superior performance with an accuracy of 87.93 ± 3.77%, sensitivity of 82.09 ± 4.03%, specificity of 91.89 ± 7.62%, and an AUROC of 0.87 ± 0.05. Compared to the image-only TCA-Net and multimodal fusion baseline, the IC-TCA-Net showed statistically significant improvements in accuracy and specificity (paired t-test, p < 0.05). Ablation analysis demonstrated that the incorporation of the dimensionality transformation module yielded substantial performance gains, significantly improving accuracy (p = 0.002) and AUROC (p = 0.010), highlighting its critical role in effective multimodal fusion. SHAP analysis quantitatively demonstrated that the dimensionality transformation module successfully mitigated modality imbalance, facilitating a synergistic contribution from both imaging and clinical data. Our findings demonstrate that addressing dimensional disparity between heterogeneous modalities is essential for mitigating modality imbalance and achieving genuine multimodal synergy. To overcome this challenge, IC-TCA-Net incorporates a dimensionality transformation module that enables effective fusion of clinical and imaging data. This approach achieves superior predictive performance for 2-year NSCLC recurrence while maintaining a parsimonious architecture suitable for real-world clinical applications.

Supplementary information: The online version contains supplementary material available at https://doi.org/10.1007/s13534-026-00588-7.