Hepatocellular carcinoma (HCC) exhibits substantial interpatient heterogeneity, leading to markedly variable outcomes and survival even among patients with similar stages and imaging phenotypes. Mainstream staging systems remain suboptimal, whereas pathology-dependent factors and high-cost genomic assays are neither scalable nor timely for clinical decision-making. Existing algorithms provide coarse risk stratification or static binary predictions, failing to capture the time-varying risk of death. We developed and externally validated TEMPO-HCC, a multimodal deep survival model with hierarchical interpretability, to estimate individualized overall survival risk trajectories after curative-intent resection. We curated a six-center cohort of 1475 patients and integrated multiphasic MRI, postoperative H&E whole-slide images, and perioperative predictors. A discrete-time survival head generated probabilities at 1, 2, 3, and 5 years after surgery. TEMPO-HCC outperformed unimodal models and guideline-based staging systems, achieving C-index of 0.751 and time-dependent AUCs of 0.836, 0.781, 0.812, and 0.680 at 12, 24, 36, and 60 months, respectively, in the external validation cohort. TEMPO-HCC represents a paradigm shift from static binary classification toward clinically actionable temporal risk prediction. Its hierarchical interpretability provides an auditable evidence chain linking macro-scale radiologic phenotypes to micro-scale histopathologic patterns. Importantly, augmenting guideline staging with TEMPO-HCC improved discrimination, enabling personalized postoperative surveillance and risk-adapted clinical management.
Fan Li, Huancheng Yang, Ruishan Liu et al.· npj Digital Medicine· 0 citations
Our study developed a multiomics-driven transformer model that combines mammography, MRI, transcriptomic and proteomic data to noninvasively predict axillary lymph node (ALN) metastasis in breast cancer. A total of 2105 patients from 10 institutions were included for model training and validation. The model achieved an AUC of 0.939 in the training cohort (n = 658) and 0.830-0.867 across three independent validation cohorts (n = 282, 971 and 194, respectively), outperforming conventional ultrasound examination. Grad-CAM visualizations highlighted the tumor edges and surrounding tissue, consistent with clinical and pathological findings. In a cohort of 194 patients, multiomics analyses linked the model output to gene and protein signatures involved in immune modulation, cytoskeletal remodeling, and epithelial-to-mesenchymal transition. Critically, the major enriched pathways identified through model-stratified analysis were independently replicated in a parallel non-model-driven analysis using ALN status, demonstrating that these signatures reflect tumor biology. Network analysis revealed gene clusters related to DNA replication and immune pathways, providing biological insights into the model's decisions. These findings suggest that the stacking model holds promise as a noninvasive decision-support tool that may complement, rather than replace, current clinical staging practices. However, integration into clinical workflows requires prospective validation.
Xiaodong Liu, Fan Li, Ye Xiang et al.· npj Precision Oncology· 0 citations
Structured reporting converts free-text radiology narratives into queryable data keys, facilitating cohort assembly, longitudinal tracking, and training label generation for medical AI. The prevailing paradigm follows a two-stage pipeline: (1) constructing a reporting template, (2) extracting information to populate it. While the extraction stage has benefited from advances in large language models (LLMs), template construction remains a manual bottleneck relying on labor-intensive expert consensus that is static, difficult to scale, and may fail to capture real-world reporting diversity. We address this limitation with ASTAR, an LLM-based framework for Automated induction of STAndardized radiology Reporting templates from large-scale clinical free-text corpora. Extensive experiments on 4,215 fetal brain MRI reports from multiple centers demonstrate that, in this reporting scenario, the ASTAR-induced template surpasses two expert-curated templates across template coverage, information fidelity, diagnostic fidelity, and expert-rated usability, reducing template development from weeks of committee deliberation to hours of automated processing.
