A MULTISCALE MULTI-VIEW CONVOLUTIONAL NEURAL NETWORK WITH ATTENTION FUSION FOR OSTEOSARCOMA PROGNOSIS PREDICTION

Osteosarcoma is a highly malignant bone tumor with a high risk of metastasis, and accurate prognosis prediction for metastasis is of great significance for clinical treatment decision-making. Traditional prediction approaches often rely on single-omics data, limiting their ability to capture the complex associations across different omics layers. To address this limitation, we propose a novel metastasis prognosis prediction model for osteosarcoma, termed Multiscale Multi-view Convolutional Neural Network with Attention Fusion (M 2 ANet), which integrates a multi-view convolutional network with an attention fusion mechanism. Specifically, we collect three types of omics data — DNA methylation, copy number variation (CNV), and Ribonucleic Acid sequencing (RNA-Seq) — from the TARGET-OS database, and perform a series of preprocessing operations. Principal component analysis (PCA) is applied to each omics type to construct multiple distinct feature views. Subsequently, a multi-scale convolutional module is employed to extract fine-grained and global features at different scales for each view, enhancing the model’s capacity to capture hierarchical biological information. Finally, an attention fusion module is introduced to dynamically assign weights across views, effectively emphasizing informative features and suppressing redundant ones, thereby improving predictive performance. By comprehensively integrating multi-source omics features, the proposed M 2 ANet model significantly improves the accuracy of metastasis prognosis prediction. Experimental results show that our model beats baseline approaches on several evaluation measures and has a high potential for clinical application.