Abstract:

[Objective] This study aims to optimize the computational performance of the Morphgene framework to address inefficiencies in processing large-scale pathology images and multi-omics data for non-small cell lung cancer (NSCLC) subtyping. [Methods] Comprehensive optimizations are applied to the framework’s pathology image patch processing, feature extraction, and K-means clustering modules through CPU thread pool scheduling, PyTorch tensor programming, GPU acceleration, and deep learning inference optimization techniques. Experiments conducted on NSCLC samples from the TCGA database validate the effectiveness of these optimizations and the framework’s subtyping performance. [Results] The optimized framework achieved over a 10-fold speedup in large-scale data processing while maintaining high subtyping accuracy. It successfully identified prognostically relevant subtypes, providing strong support for personalized treatments and survival predictions. [Limitations] The current optimizations are designed for specific file formats and patch sizes, requiring further research to adapt to smaller files or larger patches. [Conclusions] The GPU acceleration strategy significantly improves the computational efficiency of the Morphgene framework, making it a robust tool for NSCLC subtyping in precision medicine. Future work will focus on enhancing its multi-modal data integration and adaptability to broaden its clinical applications.

Key words: GPU, non-small cell lung cancer (NSCLC), multi-omics data integration, digital pathology image analysis, precision oncology