AI Reads CT Scans and Blood Tests to Spot Elusive Gut Tumors Before Surgery

A team of researchers in China has built a machine learning system that can identify gastrointestinal stromal tumors, one of the most easily missed cancers of the digestive tract, by combining the subtle shapes hidden inside CT scans with routine blood test results. The study, published in Holistic Integrative Oncology, reports that the best models achieved an area under the receiver operating characteristic curve of roughly 0.96 on an independent test set, a level of accuracy that, if confirmed in larger trials, could give surgeons a powerful non-invasive tool for deciding how to treat suspicious masses before a single incision is made.

Gastrointestinal stromal tumors, usually abbreviated GIST, are the most common mesenchymal tumors of the digestive tract, arising not from the lining of the gut but from the connective tissue within its walls. Although they represent only 0.1 to 3 percent of gastrointestinal malignancies, they carry a real malignant potential and a high risk of recurrence, which makes early detection a matter of life and death. The clinical problem is that GISTs are notoriously non-specific in their presentation. Patients may complain of vague abdominal discomfort, bleeding, or nothing at all, and by the time the tumor is found, roughly half of patients already have metastases at initial diagnosis, having lost the window for curative surgery.

The diagnostic toolkit available today has significant gaps. Endoscopy and conventional CT often fail to detect small GISTs and provide limited insight into their biological behavior. Pathological examination remains the gold standard, but obtaining tissue requires an invasive biopsy that is time-consuming and carries risks, including the theoretical possibility of tumor seeding along the needle track. What clinicians need is a way to raise or lower their suspicion of GIST using data that can be gathered without touching the patient, and that is precisely the niche the new study set out to fill.

The research team, led by Jing Yang and Hui Cai of Gansu Provincial Hospital and collaborators at Southern Medical University and Gansu University of Chinese Medicine, assembled a retrospective cohort of 317 patients treated between January 2017 and December 2024. Of these, 115 had GIST confirmed by postoperative pathology, while the remaining 202 formed a control group with other digestive tract diseases, including 93 cases of gastric cancer, 43 of colorectal cancer, 7 of duodenal cancer, 36 of gastric polyps, and 23 of gastric ulcer. The patients, whose ages ranged from 29 to 81 with a mean of about 60 years, were randomly divided into training, validation, and test sets in a 7:2:1 ratio, giving the models 221 cases to learn from, 64 to tune against, and 32 completely unseen cases for final evaluation.

On the blood side, the researchers mined an unusually comprehensive panel of 42 fasting indicators drawn from the hospital’s digital medical records, spanning 22 routine hematological parameters such as hemoglobin and platelet counts, 12 biochemical measures including liver enzymes and bilirubin fractions, and 8 tumor markers such as CEA, CA199, and the pepsinogens. Univariate analysis initially flagged 11 indicators that differed significantly between GIST and non-GIST patients, and multivariate logistic regression then distilled these down to five independent predictors: direct bilirubin (DBIL), alkaline phosphatase (ALP), CEA, CA199, and the ratio of pepsinogen I to pepsinogen II. Intriguingly, GIST patients tended to show higher DBIL and PG I/PG II values but lower CEA and CA199 than patients with other digestive diseases, a pattern the authors attribute to the distinct biology of stromal tumors compared with epithelial cancers, and possibly to the tendency of gastric GISTs to induce atrophy of the surrounding gastric mucosa.

On the imaging side, the team applied radiomics, a technique that converts medical images into hundreds of quantitative measurements invisible to the human eye. Two radiologists with more than five years of abdominal imaging experience, blinded to the pathological results, manually outlined the tumor region layer by layer on portal-phase contrast-enhanced CT images using 3D Slicer software, with a senior physician of over ten years’ experience arbitrating disagreements. After resampling the images to a uniform 1 cubic millimeter voxel size and applying Z-score normalization, the pipeline extracted 851 features, including first-order statistics, shape descriptors, texture matrices such as GLCM and GLDM, and higher-order wavelet transform features. Features that the two physicians delineated inconsistently, measured by an intraclass correlation coefficient threshold of 0.75, were discarded, and LASSO regression with five-fold cross-validation then compressed the remaining candidates to just four: Sphericity, RunVariance.7, SurfaceVolumeRatio, and LongRunEmphasis.7.

Six machine learning algorithms, namely logistic regression, k-nearest neighbors, random forest, decision tree, LightGBM, and neural networks, were then trained to build three kinds of models: one using only clinical blood features, one using only radiomic features, and one combining both. The results were striking. On the independent test set, the LightGBM radiomics model achieved the highest AUC of 0.964, while the LightGBM combined model came in essentially tied at 0.959. Both dramatically outperformed the clinical model, which topped out at an AUC of 0.745, confirming that blood indicators alone, however convenient, are not sufficient for accurate GIST diagnosis. The authors also noted honest caveats about their own data: the decision tree and LightGBM models posted perfect AUCs of 1.000 on the training set, a classic signature of overfitting, and their performance declined to varying degrees on validation and test data, a reminder that small test sets can make performance estimates jittery.

What elevates this study above many similar machine learning papers is its insistence on interpretability. Using SHAP values, a game-theoretic method that quantifies each feature’s average marginal contribution to individual predictions, the researchers found that the radiomic feature Sphericity was by far the most influential driver of the combined model’s decisions, with a mean absolute SHAP value of 2.68, followed by SurfaceVolumeRatio, while the five blood indicators contributed comparatively little, with DBIL the smallest. High sphericity values pushed predictions toward GIST, which makes biological sense: stromal tumors often grow as well-circumscribed, expansile, roughly spherical masses. Yet a forest plot of standardized coefficients told a different story, with CEA showing the largest absolute coefficient at minus 0.315. The authors argue that this divergence is itself instructive, because SHAP values capture complex feature interactions in specific predictions while regression coefficients measure net effects after controlling for other variables, and they suggest that combining multiple interpretation methods guards against over-reading any single importance metric.

Clinical utility was assessed with decision curve analysis, which weighs the benefits of acting on a model’s prediction against the harms of false positives across a range of decision thresholds. Across the clinically relevant range of 0.1 to 0.7, both the LightGBM radiomics model and the combined model delivered substantial net benefit, with their curves nearly overlapping and sitting far above the treat-all and treat-none reference strategies. At a typical decision threshold of 0.4, using either model instead of treating no one would yield roughly 35 additional net-benefit cases per 100 patients, a figure that translates statistical performance into something a clinician can weigh at the bedside.

The study has limits the authors acknowledge candidly. It is a single-center retrospective analysis with inherent selection bias, the test set of 32 patients is small, and, perhaps most importantly, the control group did not include the mesenchymal tumors that are hardest to distinguish from GIST, such as leiomyoma and schwannoma, which means the reported specificity may be optimistic. Manual delineation of tumor regions, however standardized, can never be entirely free of subjectivity. The team plans multi-center, large-sample prospective studies with a broader spectrum of differential diagnoses to validate and refine the approach. Even so, the core message stands: quantitative shape information buried in ordinary contrast-enhanced CT scans, fused with a handful of routine blood values and interpreted through transparent machine learning, can discriminate GIST from its mimics with an accuracy that approaches pathological confirmation, without a single invasive procedure. If future trials bear this out, the humble preoperative CT could become far more than a picture; it could become a diagnosis.

Subject of Research: Machine learning diagnosis of gastrointestinal stromal tumors using CT radiomics and blood indicators

Article Title: Machine learning model integrating CT radiomics and clinical blood indicators for auxiliary diagnosis of gastrointestinal stromal tumors

Article References: Machine learning model integrating CT radiomics and clinical blood indicators for auxiliary diagnosis of gastrointestinal stromal tumors. (n.d.). https://doi.org/10.1007/s44178-026-00267-8

Image Credits: AI Generated

DOI: 10.1007/s44178-026-00267-8

Keywords: gastrointestinal stromal tumors, radiomics, machine learning, LightGBM, contrast-enhanced CT, blood biomarkers, SHAP values, LASSO regression, diagnostic model, decision curve analysis, medical imaging, artificial intelligence