AI blends CT imaging and gene data to predict gastric cancer survival

Gastric cancer remains one of the world’s most lethal malignancies, and predicting which patients will survive has long depended on a narrow set of clinicopathological measures such as tumor stage, grade, and lymph node involvement. Now, a team of researchers in China has demonstrated that a machine learning model can classify the survival status of gastric cancer patients by fusing three very different kinds of data: quantitative features extracted from routine CT scans, patterns of gene co-expression derived from tumor transcriptomics, and standard clinical variables. The study, published in BMC Medical Imaging, offers a glimpse of how multimodal artificial intelligence could sharpen prognostic prediction in oncology.

The research, led by Xinxin Zhang of Jiading District Central Hospital and Jianguang Jia of Bengbu Medical University, drew on publicly available data from The Cancer Genome Atlas Stomach Adenocarcinoma collection (TCGA-STAD) and The Cancer Imaging Archive (TCIA). From these resources, the team assembled a matched cohort of 46 gastric cancer patients for whom CT imaging, transcriptomic profiles, and clinical records were all available. While modest in size, the cohort allowed the investigators to build and test an integrated model that links what a tumor looks like on a scanner with what its genes are doing at the molecular level.

The imaging side of the pipeline relied on radiomics, a technique that converts medical images into hundreds of quantitative descriptors. Using the open-source PyRadiomics toolkit, the researchers extracted 116 features from CT regions of interest. These included texture measures derived from gray-level co-occurrence, run-length, size-zone, dependence, and neighborhood gray-tone difference matrices, all of which capture subtle patterns of heterogeneity within the tumor that the human eye cannot reliably perceive. To avoid overfitting, the team applied least absolute shrinkage and selection operator (LASSO) based L1-regularized logistic regression, a method that shrinks irrelevant coefficients to zero, and retained eleven of the most informative radiomic features for downstream modeling.

On the genomics side, the researchers took a network-based approach rather than analyzing thousands of individual genes. After normalizing transcriptomic data with Z-score transformation, they applied Weighted Gene Co-expression Network Analysis, or WGCNA, a widely used method that clusters genes with correlated expression patterns into modules. The analysis identified seven such transcriptomic modules, each summarized by a single value known as a module eigengene, which is essentially the first principal component of the module’s expression profile. Condensing thousands of genes into seven eigengenes provided a low-dimensional, biologically meaningful representation of tumor transcriptional behavior, sidestepping the curse of dimensionality that plagues genomic prediction models.

These radiomic features, module eigengenes, and clinical variables were then fed into a random forest classifier, an ensemble algorithm that builds many decision trees on bootstrapped samples of the data and averages their votes. Random forests are prized in biomedical machine learning for their robustness to noise, their tolerance of mixed data types, and their resistance to overfitting on small datasets, properties well suited to a cohort of this scale. The cohort was split into a training and internal-validation set of 37 patients and a held-out evaluation set of 9 patients, with stratified five-fold cross-validation performed within the training set to estimate performance more reliably.

The results were encouraging for an exploratory study of this size. In cross-validation, the random forest model achieved an area under the receiver operating characteristic curve of 0.85, with a 95 percent confidence interval ranging from 0.68 to 0.97. On the held-out evaluation set, the model reached an accuracy of 0.78, with precision and recall for the survival class each at 0.80, and a weighted F1 score of 0.78. An AUC of 0.85 suggests the model distinguishes survivors from non-survivors considerably better than chance, while the consistency between cross-validation and held-out performance hints that the multimodal signal is not merely an artifact of the training data.

The appeal of the approach lies in its multimodality. CT radiomics capture the physical phenotype of the tumor, including its texture, shape, and internal heterogeneity, which are influenced by factors such as cell density, necrosis, and angiogenesis. Transcriptomic module eigengenes reflect the underlying molecular machinery, from immune infiltration to proliferative signaling. Clinical variables anchor both in the patient’s real-world context. By integrating all three, the model can, in principle, detect survival-relevant signals that no single data type contains on its own, a principle increasingly recognized across precision oncology.

The authors are careful to frame the work as exploratory. With only 46 patients, the confidence intervals are wide, and the held-out set of nine patients is far too small to serve as definitive external validation. The study is also retrospective, relying on publicly available de-identified data rather than a prospectively enrolled cohort, and the researchers note that no additional ethical approval was required for this secondary analysis. Larger, independent, and ideally multi-center cohorts will be needed to determine whether the model generalizes beyond this dataset, whether the same radiomic and transcriptomic features remain predictive across scanners and populations, and whether the approach outperforms existing clinicopathological risk stratification.

Still, the study adds to a rapidly growing body of literature showing that routine clinical imaging, ordinarily used for diagnosis and staging, contains a wealth of quantitative prognostic information waiting to be unlocked. Because CT scans are already obtained as part of standard gastric cancer workups, a validated radiomics-based model could eventually be deployed at little additional cost to the patient, potentially flagging high-risk individuals who might benefit from more aggressive treatment or closer surveillance. Coupling such models with genomic modules could further point clinicians toward the biological drivers behind a poor imaging phenotype.

The work was supported in part by the 2024 National Clinical Key Specialty Construction Project and several Shanghai Jiading District research funds. As artificial intelligence continues to seep into radiology and oncology, studies like this one illustrate both the promise and the discipline required: sophisticated multimodal models can extract striking predictive signals from small cohorts, but translating them into clinical tools will demand the kind of rigorous, large-scale validation that only broader collaborations can provide.