Multi-omics deep learning improves FDG PET-CT-based long-term prognostication of breast cancer

[18F] fluorodeoxyglucose positron emission tomography - computed tomography (FDG PET-CT) is increasingly used for staging of breast cancer in the primary and recurrent setting, as well as in evaluating treatment response and in follow-up. Quantitative parameters derived from the primary tumor, even in non-metastatic patients (i.e., without distant metastases but possibly with nodal involvement), have shown prognostic value. Beyond visual interpretation, quantitative evaluations may improve diagnostic accuracy and reproducibility. However, current studies often rely on predefined parameters such as maximum standardized uptake value (SUVmax), metabolic tumor volume (MTV), and total lesion glycolysis (TLG), which may overlook the high-dimensional patterns inherent in FDG PET-CT. To address this, we conducted a deep-learning-based analysis of FDG PET-CT from a large retrospective cohort of non-metastatic breast cancer patients, evaluating prognostic value from multiple perspectives. To improve patient prognosis and risk stratification, we developed a multi-omics prognostic stratification (MOPS) model that integrates clinical data, FDG PET-CT, and corresponding medical reports using CMA and transformer-based architectures to predict overall survival (OS) and disease-free survival (DFS). To support clinical applicability, we incorporated interpretability into the model, providing causal explanations, visualization-based insights, and semantic interpretations to help clinicians understand and apply the predictions transparently. The MOPS model markedly improves survival prediction, outperforming single-omics models, TN staging, and molecular subtyping, with C-index values of 0.75 (95% CI: 0.69-0.81) for OS and 0.71 (95% CI: 0.65-0.77) for DFS.