• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Monday, November 16, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Cancer

Preoperative CT Machine Learning Model Predicts Gastrointestinal Stromal Tumor Recurrence

by
November 11, 2026
in Cancer, Health
Reading Time: 7 mins read
0
Preoperative CT Machine Learning Model Predicts Gastrointestinal Stromal Tumor Recurrence

Preoperative CT Machine Learning Model Predicts Gastrointestinal Stromal Tumor Recurrence

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

A multicenter retrospective investigation has demonstrated that a machine learning model analyzing preoperative contrast-enhanced computed tomography scans can noninvasively predict recurrence-free survival in patients diagnosed with gastrointestinal stromal tumors who undergo radical surgical resection followed by adjuvant targeted therapy. Published in the journal Cancer Imaging by a collaborative team of clinical investigators, the study evaluated whether high-dimensional computational features extracted from standard preoperative cross-sectional imaging could capture microscopic biological variations and clinical behavior that conventional pathological staging systems frequently miss. Gastrointestinal stromal tumors represent the most prevalent mesenchymal neoplasms of the human digestive tract, exhibiting a remarkably heterogeneous clinical course ranging from indolent localized nodules to aggressive malignancies prone to widespread metastatic dissemination. Although curative-intent surgical intervention remains the principal therapeutic modality for primary localized disease, postoperative recurrence presents a persistent clinical challenge that historically affects up to half of all high-risk individuals even following complete surgical clearance with negative resection margins. The emergence of selective tyrosine kinase inhibitors, most notably imatinib mesylate, has established postoperative targeted maintenance as a foundational component of modern clinical care, yet optimizing individual patient management requires more dependable prognostic indicators.

Clinical decision-making regarding risk stratification and the postoperative allocation of targeted tyrosine kinase inhibitor maintenance currently relies on established macroscopic and histopathological assessment tools, primarily the Armed Forces Institute of Pathology criteria and the modified National Institutes of Health classification system. While these traditional classification frameworks have guided clinical management for decades, the study investigators noted that they exhibit notable diagnostic limitations when applied in routine practice. Conventional systems demonstrate suboptimal prognostic accuracy in differentiating outcomes for intermediate-risk cases, a shortfall that often precipitates the clinical overtreatment of biologically indolent tumors or, conversely, the insufficient therapeutic management of aggressive subtypes. Furthermore, traditional risk metrics are established almost exclusively through postoperative histopathological assessments of excised surgical specimens, depending heavily on discrete parameters such as macroscopic tumor dimensions, anatomical organ location, and microscopic mitotic counts across representative tissue sections. Because these pathological evaluations rely on small histological samples, they remain inherently susceptible to biopsy sampling bias and significant inter-observer variability, while offering no diagnostic utility prior to surgery when critical choices regarding neoadjuvant therapy and surgical approach are formulated.

To overcome these clinical shortcomings, the research team designed a dual-center retrospective cohort study to construct and externally validate a specialized computational prognostic tool designated as Modelradiomic. The investigators screened consecutive surgical candidates across two major institutional medical centers, enrolling patients who underwent radical resection for histopathologically verified primary intermediate-risk or high-risk gastrointestinal stromal tumors. Center I provided patient records spanning from September 2011 through January 2023, while Center II supplied records spanning from January 2018 through January 2023. Strict study eligibility criteria mandated that all participating subjects had undergone contrast-enhanced abdominal computed tomography within thirty days prior to their surgical resection, presented with confirmed intermediate-risk or high-risk disease according to the 2008 modified National Institutes of Health criteria, and adhered to postoperative adjuvant imatinib therapy. The investigators excluded candidates who had undergone preoperative neoadjuvant treatment, those presenting with baseline distant metastases or tumor rupture prior to surgery, cases compromised by substantial motion artifacts or incomplete anatomical tumor coverage on computed tomography scans, and individuals lost to postoperative clinical follow-up.

The final study cohort comprised 192 adult patients, including 99 male and 93 female participants, who were systematically allocated across three distinct experimental groups to build, tune, and test the machine learning architecture. Patients treated at Center I were randomly divided according to an eight-to-two ratio into a primary training cohort of 113 individuals with a mean age of 53.6 years and a testing cohort of 29 individuals with a mean age of 53.9 years. An independent external validation cohort was established from Center II, consisting of 50 patients with a mean age of 57.6 years, ensuring that the predictive algorithm could be tested on entirely separate clinical datasets derived from different institutional environments. Within the overall pooled population of 192 patients, the mean maximal tumor diameter measured 7.1 centimeters with a wide span ranging from 2.0 to 24.4 centimeters, and 14.6 percent of all evaluated neoplasms exceeded 10.0 centimeters in largest dimension. Anatomically, 63.5 percent of the primary neoplasms originated in the stomach, while 32.3 percent were situated within the small intestine, and histopathological risk profiling classified 27.1 percent of patients as intermediate risk and 72.9 percent as high risk.

Acquiring dependable and standardized radiomic data across diverse clinical environments required rigorous imaging preparation and computational normalization protocols across the participating hospitals. Patients underwent a standardized bowel preparation protocol prior to computed tomography acquisition, consisting of a low-residue diet, a mandatory ten-hour fasting interval, and the ingestion of 600 to 1000 milliliters of water roughly 40 to 60 minutes before scanning to accomplish sufficient luminal gastric distension. Cross-sectional examinations were conducted across eight separate multi-detector computed tomography scanners, necessitating mathematical harmonization to ensure structural feature stability across varying hardware specifications. Radiomic processing focused specifically on portal venous phase contrast acquisitions because of their superior visual depiction of the boundary contrast between soft-tissue tumor margins and surrounding parenchymal structures. To eliminate dimensional inconsistencies caused by differing slice thicknesses and detector geometries across machines, all digital scan volumes were standardized through three-dimensional linear interpolation into isotropic voxel dimensions measuring exactly 0.75 by 0.75 by 5 millimeters prior to feature extraction.

Volumetric segmentation of each neoplasm was performed via three-dimensional manual region-of-interest boundary delineation across portal venous phase transverse image slices utilizing ITK-SNAP software. To mitigate partial volume averaging artifacts and avoid numerical contamination from adjacent anatomical structures or peritumoral adipose tissue, the delineations incorporated an intentional one-millimeter inward boundary erosion from the visible peripheral margins of the mass. Primary volumetric segmentation was carried out by an experienced radiologist, and an independent reassessment of 25 randomly chosen patient cases was executed by a second radiologist to calculate inter-observer reproducibility. This validation produced an exceptional Dice similarity coefficient of 0.967 with a narrow 95 percent confidence interval spanning 0.949 to 0.983, confirming high spatial consistency across manual delineations. Following segmentation, the team extracted 105 high-dimensional quantitative candidate features using the open-source PyRadiomics computational framework, capturing first-order statistical distributions, geometric shape parameters, and second-order texture matrices including gray-level co-occurrence, run-length, size-zone, and neighboring gray-tone difference matrices designed to mathematically characterize intratumoral heterogeneity.

To avoid mathematical overfitting and isolate the most dependable prognostic indicators from the expansive pool of image biomarkers, the investigators executed a structured, multi-tier dimensionality reduction pipeline within the primary training dataset. Variables displaying zero variance across the patient population were removed first, after which univariate Cox proportional hazards regression was conducted as an initial statistical filter, preserving only individual metrics achieving significant association with recurrence-free survival at a significance threshold below 0.05. Pearson correlation analysis was then implemented across remaining candidates to identify and eliminate collinear pairs exhibiting correlation coefficients greater than 0.85, retaining only the feature showing stronger prognostic correlation. The most informative surviving parameters were subsequently isolated via least absolute shrinkage and selection operator Cox regression, selecting the optimal tuning penalty parameter through tenfold cross-validation guided by the minimum error criterion. This rigorous algorithmic pruning ultimately identified four robust radiomic features: first-order kurtosis, gray-level non-uniformity derived from the size-zone matrix, busyness from the neighboring gray-tone difference matrix, and coarseness from the same spatial matrix.

These four selected imaging biomarkers were integrated into an ensemble random survival forest machine learning model configured with a tree depth of two, designed to forecast individual recurrence-free survival timeframes. From a pathophysiological perspective, the investigators highlighted that elevated values for kurtosis and gray-level non-uniformity reflect an erratic and disorderly distribution of computed tomography voxel intensities across the tumor volume. In gastrointestinal stromal tumor biology, this uneven density profile frequently correlates with internal areas of intratumoral hemorrhage, cystic degeneration, or central ischemic necrosis, all of which represent established hallmarks of aggressive cellular proliferation and high-grade malignancy. Concurrently, neighboring gray-tone difference matrix busyness and coarseness measure the spatial rate of voxel intensity transitions between adjacent microstructures. A high busyness value mathematically reflects rapid spatial transitions characteristic of a dense, complex stromal architecture and accelerated microvascular neoangiogenesis, providing the biological foundation and vascular scaffolding that facilitate early postoperative recurrence and distant hematogenous dissemination toward the liver and peritoneal surfaces.

Across all evaluated patient groups during a median post-surgical follow-up period of 65.0 months, recurrence developed in 24.0 percent of the total study population, representing 46 individuals with a median recurrence-free survival time of 57.5 months. The predominant anatomical sites of recurrent disease were the mesentery and peritoneal cavity in 45.7 percent of cases, followed by hepatic metastases in 37.0 percent, isolated surgical site recurrences in 28.3 percent, and multi-site metastases documented in five patients. When evaluating overall predictive capability, the machine learning model achieved concordance index values of 0.763 in the primary training cohort, 0.746 in the testing cohort, and 0.678 in the external validation cohort. These computational scores consistently surpassed the prognostic accuracy achieved by the conventional Armed Forces Institute of Pathology index, which yielded lower concordance index values of 0.660, 0.616, and 0.539 across the respective datasets, representing statistically significant performance differences in both the training cohort at p equal to 0.010 and the external validation cohort at p equal to 0.040.

Time-dependent receiver operating characteristic curve evaluations further confirmed the superior prognostic accuracy of the computational radiomics model, particularly over prolonged follow-up periods following surgical resection. At the three-year postoperative benchmark, the machine learning tool achieved area under the curve values ranging between 0.744 and 0.790 across cohorts, compared to values between 0.497 and 0.649 for the conventional pathology index. Kaplan-Meier survival curves revealed that patients stratified into the high radiomic score group exhibited markedly shortened recurrence-free survival and significantly elevated recurrence rates compared to those within the low-score group at p below 0.001.

The study authors highlighted that this preoperative machine learning framework is designed to complement rather than replace established postoperative histological risk grading systems, providing vital noninvasive diagnostic insight prior to surgery. By evaluating the entire three-dimensional volume of the tumor on standard preoperative computed tomography scans, the model captures global spatial heterogeneity and bypasses the tissue sampling errors that inherently constrain standard histological sectioning. This noninvasive approach offers immediate preoperative utility by identifying aggressive biological behavior in patients who may benefit from neoadjuvant imatinib regimens or tailored surgical margins, while resolving diagnostic ambiguity in borderline intermediate-risk presentations.

Despite these encouraging findings, the investigators detailed several important study limitations that warrant careful consideration before clinical implementation can be considered. Although multicenter in design, the total patient cohort of 192 individuals remains modest for machine learning applications, which could introduce vulnerabilities related to sampling variance despite the use of rigorous feature selection and external validation. Additionally, the retrospective dataset lacked granular genomic information regarding specific oncogenic driver mutations, such as distinctions between KIT exon 11 mutations and PDGFRA alterations, as well as exact dosing schedules and durations of postoperative imatinib therapy, which are known to influence long-term clinical trajectories. The required minimum follow-up duration of 24 months also meant that certain patients were still receiving active tyrosine kinase inhibitor therapy at their last recorded assessment, restricting the capacity to evaluate delayed disease rebound following treatment cessation. Finally, because the study population consisted entirely of Chinese patients imaged across eight specific scanner models, the authors emphasized that prospective multi-ethnic investigations integrating radiogenomic profiling are essential to confirm the generalizability of these findings across broader global populations.

Subject of Research: Machine learning based on preoperative CT for noninvasive prediction of recurrence-free survival in gastrointestinal stromal tumors

Article Title: Machine learning based on preoperative CT for noninvasive prediction of recurrence-free survival in gastrointestinal stromal tumors

Article References: Zhao, L., Zhao, L., Yin, X., Yuan, E., Gu, Y., Zheng, X., Wu, B., Yin, Y., & Liu, X. (2026). Machine learning based on preoperative CT for noninvasive prediction of recurrence-free survival in gastrointestinal stromal tumors. Cancer Imaging, 26(1), Article 119. https://doi.org/10.1186/s40644-026-01103-y

Image Credits: AI Generated

DOI: 10.1186/s40644-026-01103-y

Keywords: Machine, learning, based, preoperative, noninvasive, prediction, recurrence-free, survival, gastrointestinal, stromal, tumors

News Source: Nathaniel Bowman. (November 11, 2026). Preoperative CT Machine Learning Model Predicts Gastrointestinal Stromal Tumor Recurrence. Scienmag.

Tags: Basedgastrointestinallearningmachinenoninvasivepredictionpreoperativerecurrence-freestromalsurvivaltumors
Share12Tweet7Share2ShareShareShare1

Related Posts

Sports Ethicists Urge Strict Limits on Performance Enhancing Swimming and Running Technologies

Sports Ethicists Urge Strict Limits on Performance Enhancing Swimming and Running Technologies

November 16, 2026
Case Study Links Antipsychotic Drug Clozapine to Rare Colon Eosinophilic Cryptitis

Case Study Links Antipsychotic Drug Clozapine to Rare Colon Eosinophilic Cryptitis

November 16, 2026

Over Seventy Percent of Neuropsychiatric Outpatients Show Low Treatment Adherence

November 16, 2026

Mapping Dental Clinic Clusters Using Mobile Geospatial Technology in India

November 16, 2026

POPULAR NEWS

  • Emerald Ash Borer Infestation Triggers Divergent Root Bacterial Shifts in Ash

    Emerald Ash Borer Infestation Triggers Divergent Root Bacterial Shifts in Ash

    29 shares
    Share 12 Tweet 7
  • Lespedeza cyrtobotrya Stem Extracts Inhibit Monoamine Oxidase Enzymes In Vitro

    29 shares
    Share 12 Tweet 7
  • Sports Ethicists Urge Strict Limits on Performance Enhancing Swimming and Running Technologies

    29 shares
    Share 12 Tweet 7
  • Case Study Links Antipsychotic Drug Clozapine to Rare Colon Eosinophilic Cryptitis

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Emerald Ash Borer Infestation Triggers Divergent Root Bacterial Shifts in Ash

Lespedeza cyrtobotrya Stem Extracts Inhibit Monoamine Oxidase Enzymes In Vitro

Sports Ethicists Urge Strict Limits on Performance Enhancing Swimming and Running Technologies

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 89 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.