Radiomics and Deep Learning for Non-Invasive MSI Detection in Colorectal Cancer

Abstract

Microsatellite instability (MSI) is a crucial molecular biomarker in colorectal cancer (CRC) that significantly influences prognosis, therapeutic strategies, and immunotherapy response. Conventional MSI detection relies on invasive tissue-based techniques such as polymerase chain reaction (PCR) and immunohistochemistry (IHC), which are associated with procedural risks, high costs, and potential sampling bias. With the growing availability of medical imaging and advancements in artificial intelligence (AI), non-invasive MSI prediction has emerged as a promising alternative. Radiomics enables the extraction of quantitative features from medical images, capturing tumor heterogeneity, while pre-trained deep learning models provide robust hierarchical representations through transfer learning. This survey presents a comprehensive analysis of methods that integrate radiomics with hyperparameter-tuned pre-trained models for MSI detection in CRC. It examines various architectures including convolutional neural networks, transformer-based models, and hybrid frameworks that combine handcrafted and deep features. Emphasis is placed on optimization strategies such as Bayesian optimization, grid search, and evolutionary algorithms to enhance model performance. A systematic review of 30 studies is conducted to evaluate methodological advancements, datasets, and predictive outcomes. The findings reveal that hybrid approaches consistently outperform standalone radiomics or deep learning models, achieving improved accuracy, generalization, and robustness. Hyperparameter tuning further contributes to performance stability and clinical applicability. This survey identifies key trends, challenges, and future directions for developing reliable, non-invasive MSI prediction systems suitable for real-world clinical integration.