• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Wednesday, November 18, 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 Health

Deep Learning Identifies Medulloblastoma Molecular Subgroups and Genetic Signatures Using MRI

by
November 14, 2026
in Health
Reading Time: 6 mins read
0
Deep Learning Identifies Medulloblastoma Molecular Subgroups and Genetic Signatures Using MRI

Deep Learning Identifies Medulloblastoma Molecular Subgroups and Genetic Signatures Using MRI

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

Deep learning models analyzing preoperative magnetic resonance imaging can accurately classify medulloblastoma molecular subgroups and identify high-risk genetic signatures, according to a multicenter retrospective study in the Chinese Neurosurgical Journal. The computational framework, employing sequential deep neural networks and automated tumor segmentation, stratified patients into four molecular categories and predicted actionable genetic alterations across clinical cohorts. In an external validation cohort comprising 124 patients from an independent hospital, the primary classification network achieved a median diagnostic accuracy of 77.50 percent across medulloblastoma subgroups. Furthermore, specialized secondary deep learning models predicted high-risk signatures, detecting TP53 mutations in sonic hedgehog tumors, MYC amplification in Group 3, and chromosome 11 loss in Group 4. Combining neural network outputs with patient demographics and conventional radiological features in a hybrid model raised overall diagnostic accuracy to 82.20 percent.

Medulloblastoma is the most prevalent malignant pediatric brain tumor arising in the posterior fossa, characterized by marked biological and clinical heterogeneity. Under the 2021 World Health Organization classification framework, the malignancy is divided into four molecular subgroups: wingless (WNT), sonic hedgehog (SHH), Group 3, and Group 4. Prognosis varies substantially across these subgroups, carrying direct implications for risk-adapted therapy. For instance, WNT tumors demonstrate favorable therapeutic outcomes with five-year overall survival exceeding 90 percent. Conversely, sonic hedgehog tumors exhibit prognostic heterogeneity driven by TP53 mutation status. Meanwhile, Group 3 and Group 4 tumors harboring MYC amplification or chromosome 11 loss exhibit aggressive clinical behavior and poor treatment response. Distinguishing these subgroups and identifying associated genetic variations prior to surgical intervention is vital for guiding therapeutic intensity, avoiding undertreatment in high-risk patients, and reducing over-treatment toxicities in low-risk individuals.

Although deep learning has shown success in adult gliomas for predicting biomarkers such as MGMT promoter methylation and IDH mutations, applying artificial intelligence to pediatric medulloblastoma has faced notable hurdles. Earlier medulloblastoma models achieved approximately 70 percent to 85 percent subgroup classification accuracy but were constrained by limited cohort sizes and lacked external validation on independent datasets. Furthermore, previous frameworks typically excluded high-risk genetic signatures such as TP53 mutations, MYC amplification, and chromosome 11 loss, while failing to integrate multimodal clinical and imaging parameters. Conventional radiological evaluation by neuroradiologists also introduces subjective assessment and risk of misclassification. To bridge this translational gap between noninvasive imaging and actionable molecular risk assessment, the researchers designed a dual-stage framework integrating preoperative magnetic resonance imaging with ground truth derived from DNA methylation arrays and next-generation sequencing, including whole-genome sequencing and comparative genomic hybridization.

The multicenter retrospective investigation enrolled 449 consecutive patients with newly diagnosed medulloblastoma treated between January 2015 and June 2023 across two medical institutions. The study protocol received institutional review board approval with informed consent waived due to minimal risk. The total patient population included 249 males and 200 females, with a mean age of 13.55 years and a standard deviation of 2.33 years. Age distribution analyses identified three distinct cohorts: 14 patients younger than 3 years (mean age 1.86 ± 1.06 years), 243 patients aged 3 to 18 years (mean age 11.30 ± 3.12 years), and 68 adult patients aged 18 to 49 years (mean age 28.67 ± 4.33 years). Statistical testing indicated no significant differences in age (P = 0.13) or sex distribution (P = 0.28) across the four medulloblastoma molecular subgroups.

For computational model development and internal validation, 325 patients from Beijing Tiantan Hospital formed the training cohort, while an independent external test cohort of 124 patients from Beijing Shijitan Hospital assessed generalizability. A subset of 375 patients with verified genetic signatures was selected to construct second-stage predictive models. Inclusion criteria required confirmed medulloblastoma and complete preoperative axial scans, encompassing T1-weighted, T2-weighted, and contrast-enhanced T1-weighted sequences. Exclusion criteria removed patients lacking molecular or genetic data, individuals with prior neurosurgical procedures before imaging, and scans with severe radiological artifacts. Magnetic resonance imaging was performed using 1.5-Tesla and 3.0-Tesla systems from Philips Healthcare, Siemens Healthineers, GE Healthcare, and Toshiba Medical Systems. Scans were converted into NIfTI format via FSL software, followed by spatial resampling and intensity normalization from -1 to 1 to reduce scanner variability.

To automate tumor localization, the authors implemented a U-Net architecture to generate three-dimensional tumor segmentations. The U-Net encoder-decoder structure extracts high-dimensional representations through convolutional and pooling layers, while skip connections restore precise spatial details. Developed on the 325-patient cohort (split into 80 percent training, 10 percent validation, and 10 percent testing) over 100 epochs using the Adam optimizer with cosine learning rate decay and Dice loss, the segmentation model achieved a mean Dice score of 0.91 ± 0.13. Two neuroradiologists independently reviewed the automated boundaries, with low-quality outputs re-segmented using ITK-SNAP and adjudicated by a senior neuroradiologist with 20 years of clinical experience. Concurrently, two neuroradiologists evaluated qualitative imaging characteristics across all cohorts, recording T1-weighted hyperintense foci, T2-weighted hypointense foci, and contrast enhancement levels with high inter-reader agreement, yielding kappa scores between 0.842 and 0.864 (P < 0.001).

For molecular subgroup classification, the team benchmarked four algorithms on the validation dataset: VGG, U-Net, GoogLeNet, and ResNet-50. Based on precision with 95 percent confidence intervals, VGG achieved 65.03 percent (95% CI: 47.28% to 71.66%), U-Net reached 72.00 percent (95% CI: 62.78% to 82.33%), GoogLeNet yielded 69.06 percent (95% CI: 55.43% to 79.87%), and ResNet-50 led with 78.05 percent (95% CI: 69.00% to 86.33%). ResNet-50 was therefore chosen as the classification core, named MB-CNN. Its 50-layer architecture uses residual shortcuts to facilitate deep gradient propagation. Trained with focal loss and the Adam optimizer on PyTorch, MB-CNN outputs a four-element probability vector via softmax activation. In internal validation, MB-CNN differentiated subgroups with accuracy between 74.68 percent and 77.84 percent, precision between 73.85 percent and 78.90 percent, recall from 74.69 percent to 77.81 percent, and F1 scores from 63.60 percent to 80.00 percent. On the independent external test dataset of 124 patients, MB-CNN achieved a median accuracy of 77.50 percent, ranging from 76.29 percent to 78.71 percent.

When tumors were identified as sonic hedgehog, Group 3, or Group 4, the diagnostic pipeline advanced to three specialized binary ResNet-50 models: MB-CNN_TP53, MB-CNN_MYC, and MB-CNN_Chr11. These networks analyzed raw imaging and U-Net segmentations to predict high-risk genetic signatures. In internal validation, the TP53 model in sonic hedgehog tumors achieved an area under the receiver operating characteristic curve of 0.91 (95% CI: 0.82 to 0.96) and 90.14 percent accuracy (95% CI: 87.63% to 96.42%). The MYC model in Group 3 achieved an area under the curve of 0.87 (95% CI: 0.76 to 0.94) and 84.43 percent accuracy (95% CI: 77.19% to 93.53%). The chromosome 11 loss model in Group 4 yielded an area under the curve of 0.89 (95% CI: 0.80 to 0.98) and 88.64 percent accuracy (95% CI: 73.42% to 99.44%). On the 105 external test patients with genetic data (44 sonic hedgehog, 23 Group 3, and 38 Group 4), the TP53 model achieved an area under the curve of 0.93 (95% CI: 0.85 to 0.97) and 91.27 percent accuracy; the MYC model reached an area under the curve of 0.85 (95% CI: 0.79 to 0.92) and 83.56 percent accuracy; and the chromosome 11 loss model demonstrated an area under the curve of 0.88 (95% CI: 0.79 to 0.95) and 86.79 percent accuracy.

To investigate whether multimodal integration could enhance diagnostic efficacy, the researchers constructed a baseline logistic regression model and a hybrid model incorporating MB-CNN outputs alongside clinical information and conventional imaging characteristics. Evaluated on the independent external test cohort, the logistic regression model relying solely on patient age, sex, and conventional radiological features showed limited efficacy, achieving an accuracy between 58.28 percent and 65.22 percent, precision from 53.85 percent to 66.67 percent, and an overall accuracy of 59.14 percent. In contrast, the hybrid model achieved an overall accuracy of 82.20 percent (range 78.00 percent to 86.21 percent) and precision between 75.86 percent and 88.89 percent. The hybrid model significantly outperformed the conventional logistic regression baseline (82.20 percent versus 59.14 percent, P = 0.009) and showed comparable performance to standalone MB-CNN (82.20 percent versus 77.50 percent, P = 0.105), representing a mean accuracy gain of 21.04 percent over logistic regression and an average improvement of 5.58 percent over MB-CNN alone.

Age-stratified subgroup analyses further assessed algorithmic robustness across pediatric and adult cohorts within the independent test dataset. In the juvenile cohort (aged 18 years and younger, n = 98), MB-CNN achieved classification accuracy ranging from 68.97 percent to 85.67 percent and precision of 62.50 percent to 78.33 percent, while the secondary genetic models demonstrated areas under the curve from 0.89 to 0.96 and accuracy between 83.33 percent and 92.11 percent. In the adult cohort (aged older than 18 years, n = 26), MB-CNN achieved accuracy between 66.67 percent and 81.82 percent with precision from 69.57 percent to 78.33 percent, while the genetic signature models achieved areas under the curve from 0.88 to 0.95 and accuracy between 82.61 percent and 91.30 percent. These findings confirmed that the deep learning architecture preserved predictive accuracy across developmental stages, effectively characterizing both pediatric and adult medulloblastoma presentations.

The authors noted several limitations inherent to the study design. First, the investigation was retrospective and included patients from only two medical centers, which may limit generalizability across broader clinical populations. Second, although spatial resampling and intensity normalization were applied across 1.5-Tesla and 3.0-Tesla scanners from four manufacturers, magnetic field inhomogeneities were not adjusted using N4 bias field correction, leaving residual protocol-related variations. Third, the sample size of adult medulloblastoma patients was relatively modest, particularly within less common adult subgroups like WNT and Group 3, limiting statistical power in that demographic. Finally, while noninvasive deep learning frameworks offer a promising, cost-effective tool for preoperative molecular risk assessment, tissue histopathology and genomic profiling remain essential standards for definitive clinical management. The authors emphasized that prospective multicenter trials across diverse healthcare institutions will be necessary before artificial intelligence models can be deployed in routine neuro-oncology workflows.

Subject of Research: Deep learning and hybrid modeling for medulloblastoma molecular subgrouping and genetic signature identification using MRI

Article Title: Exploring deep learning and hybrid approaches in molecular subgrouping and prognostic-related genetic signatures of medulloblastoma

Article References: Li, Y., Liu, H., Liu, Y., Li, J., Suzuki, H. H., Liu, Y., Tao, J., & Qiu, X. (2025). Exploring deep learning and hybrid approaches in molecular subgrouping and prognostic-related genetic signatures of medulloblastoma. Chinese Neurosurgical Journal, 11(1), Article 19. https://doi.org/10.1186/s41016-025-00405-7

Image Credits: AI Generated

DOI: 10.1186/s41016-025-00405-7

Keywords: Medulloblastoma, Deep Learning, ResNet-50, U-Net, Molecular Subgrouping, Preoperative MRI, TP53 Mutation, MYC Amplification, Chromosome 11 Loss, Sonic Hedgehog Subgroup, WNT Subgroup, Group 3 Medulloblastoma, Group 4 Medulloblastoma, Hybrid Model, Neuro-Oncology, External Validation, Pediatric Brain Tumors, Radiomics

News Source: Blake Davidson. (November 14, 2026). Deep Learning Identifies Medulloblastoma Molecular Subgroups and Genetic Signatures Using MRI. Scienmag.

Tags: Chromosome 11 Lossdeep learningexternal validationGroup 3 medulloblastomaGroup 4 medulloblastomahybrid modelmedulloblastomaMolecular SubgroupingMYC amplificationNeuro-oncologypediatric brain tumorsPreoperative MRIradiomicsResNet-50Sonic Hedgehog SubgroupTP53 mutationU-NetWNT Subgroup
Share12Tweet7Share2ShareShareShare1

Related Posts

Many Health Test Claims Marketed to UK Consumers Are Unsubstantiated

Many Health Test Claims Marketed to UK Consumers Are Unsubstantiated

November 18, 2026
Multimodal Surgery and Adjuvant Therapy Resolve Massive Abdominal Desmoid Tumor

Multimodal Surgery and Adjuvant Therapy Resolve Massive Abdominal Desmoid Tumor

November 18, 2026

CH25H Pathway Directs Phagocyte Clearance and Functional Recovery After Brain Haemorrhage

November 18, 2026

Early Sensory Differences Prospectively Predict Subsequent Autistic Traits in Longitudinal Cohort

November 18, 2026

POPULAR NEWS

  • Early Sensory Differences Prospectively Predict Subsequent Autistic Traits in Longitudinal Cohort

    Early Sensory Differences Prospectively Predict Subsequent Autistic Traits in Longitudinal Cohort

    29 shares
    Share 12 Tweet 7
  • High-Power Spray Coagulation Boosts Efficiency in Colorectal Underwater Endoscopic Dissection

    29 shares
    Share 12 Tweet 7
  • Robotic Latissimus Dorsi Harvest Minimizes Incisions in Complex Breast Reconstruction

    29 shares
    Share 12 Tweet 7
  • Integrated Decision Framework Cuts Costs and Speeds Supply Network Recovery

    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

Acid Activated Copper Slag Reaches 100 Megapascals in Cementless Binder Study

Many Health Test Claims Marketed to UK Consumers Are Unsubstantiated

Multimodal Surgery and Adjuvant Therapy Resolve Massive Abdominal Desmoid Tumor

Subscribe to Blog via Email

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

Join 88 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.