Research on HIFU diagnosis and treatment assistance system: Prediction and optimization of desmoid tumor treatment based on dynamic-static feature interaction parallel network

Abstract

High-intensity focused ultrasound (HIFU) ablation is recognized as an effective treatment for desmoid tumors. However, the treatment process is heavily reliant on the clinical expertise of physicians. Moreover, limitations caused by the scarcity and heterogeneity of medical data restrict the effectiveness of feature extraction and accurate modeling. To address these challenges, a medical data-driven intelligent assistance system is proposed for HIFU treatment. Initially, an improved stacked weighted random forest is employed for data augmentation, where weighting strategies, enhanced feature matrix methods, and multi-model stacking techniques are utilized to strengthen critical features. Then, the enhanced HIFU treatment data are extracted utilizing a parallel static and dynamic feature-fusing network equipped with an enhanced module. Within this network, the static branche promotes the accurate capturing of complex feature relationships and the efficient integration of key information through feature space transformation and a global context modeling mechanism for discrete patient-related features. The dynamic branch combined with a learnable activation function accurately extracts complex nonlinear relationships related to the treatment process. Finally, a cross-attention mechanism is introduced to realize the positive and negative simultaneous enhancement between static and dynamic features, fully revealing the spatio-temporal correlation between dynamic and static features, and improving the comprehensive performance of the model. Finally, the effectiveness and superiority of the proposed method are verified through simulation and experiment. The proposed method is integrated into the HIFU assisted diagnostic and treatment framework to enhance decision-making in HIFU treatment modeling.