Multi-omics fusion network for prediction of early recurrence in colorectal liver metastases

Nearly half of colorectal cancer patients develop liver metastases. While surgical removal offers a potential cure, the majority experience recurrence within two years. Accurate tools to predict the recurrence risk are lacking. This study proposes a multi-omics framework combining computed tomography, transcriptomic (RNA) sequencing, and the Clinical Risk Score (CRS) to predict the likelihood of two-year recurrence after colorectal liver metastasis (CRLM) resection. Our approach addresses undetected RNA transcripts by introducing generative adversarial imputation and leverages generative learning and transformers to manage high dimensional gene expression data. Imaging features are extracted using a foundation model alongside interpretable radiomics. Tested on a prospectively maintained dataset of 129 patients, the pipeline achieved an area under the curve of 0.75 ± 0.05, outperforming unimodal and bimodal approaches and the CRS. This assistive tool can improve risk stratification, inform patients of their expected outcomes, guide follow-up care and inspire clinical trials for post-operative treatments.