This project aims to enhance a traditional radiomics pipeline (PyRadiomics + classical machine learning) by integrating ​3D deep segmentation/representation learning​ and ​multimodal fusion​ to improve cross-center generalizability and workflow automation. The core computational tasks heavily rely on ​GPU throughput and memory capacity​ to ensure reproducibility, stability, and timely delivery. Key innovations include automated 3D segmentation using advanced architectures like nnU-Net or Swin-UNETR, deep radiomics feature extraction via self-supervised learning, and GPU-accelerated habitat clustering in latent spaces. The resource plan allocates approximately ​1,800 GPU hours​ and ​40,000–50,000 CPU core hours​ to support a dual-track strategy that combines traditional radiomics with state-of-the-art deep learning.