• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Tuesday, November 17, 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

Transfer Learning Framework Automates Paediatric Glioma Segmentation And PET Biomarker Analysis

by
November 12, 2026
in Health, Technology
Reading Time: 6 mins read
0
Transfer Learning Framework Automates Paediatric Glioma Segmentation And PET Biomarker Analysis

Transfer Learning Framework Automates Paediatric Glioma Segmentation And PET Biomarker Analysis

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

A deep learning framework leveraging transfer learning from adult brain tumour datasets has successfully automated tumour segmentation and quantitative molecular metric extraction from [18F]F-DOPA positron emission tomography scans in paediatric glioma patients. Gliomas represent the most frequent central nervous system malignancy in children, accounting for up to thirty-five percent of childhood brain tumours and presenting diverse biological behaviors. While molecular neuroimaging provides crucial insight into tumour aggressiveness, histological boundaries, and patient prognosis, conventional clinical workflows remain heavily reliant on manual tumour contouring and region-of-interest selection. These manual procedures demand substantial clinical expertise, consume extensive operator time, and introduce unavoidable inter-observer variability that hinders quantitative reproducibility across clinical centers. To overcome these fundamental operational limitations, researchers adapted an optimized three-dimensional U-Net convolutional neural network to establish an end-to-end automated analytical pipeline bridging morphological magnetic resonance imaging and molecular positron emission tomography.

In paediatric neuro-oncological practice, magnetic resonance imaging provides the primary diagnostic foundation due to its unmatched soft-tissue contrast, vascular assessment, and high anatomical resolution. The metabolic tracer substantially refines the distinction between active neoplastic infiltration and surrounding non-malignant vasogenic oedema, aids surgical biopsy targeting, and assists radiotherapy treatment planning. Furthermore, [18F]F-DOPA uptake demonstrates clinical utility in the non-invasive detection of the histone H3K27M mutation, a critical diagnostic biomarker in diffuse midline gliomas. Directly training automated segmentation networks on raw [18F]F-DOPA PET images is impeded by heterogeneous tracer uptake profiles and minimal accumulation in select high-grade paediatric cases. The investigators resolved this challenge by segmenting lesions on fluid-attenuated inversion recovery magnetic resonance imaging, using the derived morphological mask to guide subsequent tracer metric extraction across co-registered PET volumes.

The research team retrospectively evaluated one hundred and three paediatric patients presenting with newly diagnosed, treatment-naive gliomas between 2011 and 2024. All participating subjects underwent diagnostic fluid-attenuated inversion recovery magnetic resonance scans and static or dynamic [18F]F-DOPA PET acquisitions within a strict two-week window. The clinical cohort comprised fifty-four female and forty-nine male participants, presenting with a median age of eleven years. Positron emission scans were gathered using Discovery ST or Discovery MI PET/CT scanners following intravenous radiotracer administration, while magnetic resonance imaging was conducted on 1.5 Tesla or 3.0 Tesla systems using non-contrast sequences. Cases lacking histological confirmation, missing necessary MR sequences, or exhibiting severe motion artefacts were excluded. Seventy-two subjects were randomly allocated to model fine-tuning, thirty-one patients formed an independent blind test set, ninety provided complete survival tracking and histological grading data, and thirty-two had dynamic PET acquisitions.

To overcome the clinical scarcity of paediatric neuro-imaging data, the investigators implemented a transfer learning strategy using a three-dimensional U-Net derived from the nnU-Net framework. The architecture was pre-trained on the adult Brain Tumour Segmentation 2023 challenge dataset, which incorporated one thousand two hundred and fifty-one annotated adult glioma cases and an unlabelled test set of two hundred and nineteen patients across one hundred and fifty epochs with five-fold cross-validation. Adapting the network to children required modifying the architecture from multi-region segmentation to whole-tumour binary classification, reducing input and output channels from three to one. When initiating domain adaptation on the paediatric training subset, researchers froze the initial two encoder layers of the network while retraining downstream parameters.

Volumetric images underwent standardized preprocessing to match adult BraTS coordinate conventions. The SynthStrip tool executed initial skull stripping across MR volumes. The Advanced Normalization Tools suite applied affine transformations and Symmetric Normalization deformable algorithms to register fluid-attenuated inversion recovery images onto the SRI-24 anatomical atlas with isotropic one-millimeter voxels. Binary lesion annotations derived from expert-reviewed semi-automatic segmentations were resampled using nearest-neighbor interpolation, while MR volumes utilized trilinear interpolation. Volumes were aligned to the Left Posterior Superior coordinate orientation via NiBabel. To bridge the adult-paediatric domain gap and address under-represented small tumours measuring under three thousand voxels with poorly defined borders, extensive data augmentation generated twenty synthetic variations per subject across ten patients using PyTorch. Transformations incorporated axial flipping, elastic deformation, Gaussian noise, bias field shifts, and gamma corrections, expanding the paediatric fine-tuning cohort to two hundred and seventy-two volumetric series.

Evaluating fine-tuning configurations revealed key insights regarding algorithmic convergence and generalization. Models trained without data augmentation demonstrated deceptively high cross-validation metrics during initial training folds. However, their learning trajectories exhibited flat curves over one hundred and fifty epochs, indicating that model weights failed to update and adapt to paediatric cases. Consequently, unaugmented models failed when deployed on the independent blind test set of thirty-one patients, scoring below 0.80. In contrast, training on augmented cohorts acted as an effective regularizer, prompting continuous optimization and steady Dice improvements. On the blind test cohort, the augmented fluid-attenuated inversion recovery model achieved a Dice score of 0.81 ± 0.13, with fold three attaining 0.82 ± 0.11 and scoring 0.79 ± 0.15 across the full cohort. Integrating additional T1-weighted sequences degraded performance to 0.78 ± 0.18, confirming that adding multimodal channels within modest cohorts promoted overfitting.

Following model selection, the optimal architecture was embedded into an end-to-end processing pipeline to validate extracted PET indices against standard clinical methods. To evaluate the full cohort while preventing circular testing on the seventy-two fine-tuning patients, investigators introduced controlled mathematical image perturbations. These perturbations incorporated Gaussian noise with standard deviations between 0 and 0.05 alongside localized elastic deformations. Wilcoxon signed-rank tests across all one hundred and three patients confirmed that these structural perturbations caused no statistically significant alterations in extracted static or dynamic PET indices, with all comparisons yielding non-significant p-values above 0.05. The automated framework was subsequently evaluated against two reference approaches: semi-automatic region-growing segmentation executed in MRIcroGL software and manual clinical delineation using the PET VCAR tool on a GE Healthcare AW workstation seeded directly at maximal radiotracer uptake points.

Statistical validation across ninety patients demonstrated outstanding consistency for static tumour-to-striatum uptake ratios across all three segmentation approaches. Non-parametric Wilcoxon tests with Bonferroni correction showed no significant differences between automatic and semi-automatic extraction (p = 0.660), automatic and manual delineation (p = 0.342), or manual and semi-automatic methods (p = 0.639), with Spearman correlation coefficients exceeding 0.93. Conversely, the tumour-to-background ratio differed significantly when comparing manual tracing against semi-automatic and automatic pipelines (p < 0.005), yielding systematic biases of -0.252 and -0.224. Methodological inspection revealed this divergence resulted from differing background volumes: manual analysis used a spherical contralateral centrum semiovale volume, whereas automated pipelines sampled the entire contralateral white matter. Harmonizing background selection by applying the centrum semiovale across all workflows eliminated all statistical differences, confirming that automated lesion delineation was highly accurate. Furthermore, Bland-Altman analyses demonstrated that patient ranking remained stable across all comparisons, with Spearman correlation values remaining above 0.91.

The automated extraction of dynamic PET metrics was evaluated across thirty-two patients with four-dimensional acquisitions, providing exploratory evidence of kinetic agreement. Evaluated dynamic indices included Time-to-Peak, striatum slope, tumour slope calculated via linear regression from two minutes post-injection to scan conclusion, and Dynamic Slope Ratio. Striatum slope and Time-to-Peak showed no statistically significant differences between methods, with striatum slope correlations exceeding 0.90 across all comparisons. However, Bland-Altman analyses revealed systematic biases relative to manual delineation, including a Time-to-Peak bias of 2.464 for the automatic method. Manual and semi-automatic methods differed significantly for tumour slope (p = 0.008) and Dynamic Slope Ratio (p = 0.012), but automatic comparisons showed no significant differences after Bonferroni correction (p = 0.050 and p = 0.056). Kinetic curves proved highly sensitive to minor voxel-level variations due to lower dynamic PET spatial resolution. Residual biases of 0.246 for tumour slope and 0.695 for Dynamic Slope Ratio were also documented.

To establish clinical utility, automated static indices were evaluated against histological tumor grades and survival outcomes in ninety patients. Cohort classification comprised forty-nine high-grade and forty-one low-grade glioma cases. Mann-Whitney U testing revealed significant separations between tumor grades for both automated tumour-to-striatum (p < 0.05, Cohen’s D = -1.05) and tumour-to-background ratios (p < 0.05, Cohen’s D = -0.79). Furthermore, Kaplan-Meier analyses utilizing median thresholds of 1.08 for tumour-to-striatum and 1.24 for tumour-to-background demonstrated robust risk stratification. Log-rank testing yielded statistically significant differences (p < 0.05) between high-uptake and low-uptake cohorts for both overall survival and progression-free survival. These clinical correlations confirm that automated PET extraction accurately distinguishes aggressive from low-grade pediatric neoplasms and delivers reliable prognostic stratification without requiring labor-intensive manual delineation. Information regarding biological aggressiveness and patient prognosis remains central to charting therapeutic regimens in paediatric neuro-oncology, and molecular PET imaging allows non-invasive tracking across entire neoplastic volumes.

The researchers highlighted several technical and clinical limitations that frame their findings. The investigation was conducted retrospectively at a single clinical institution, constrained by the cohort size inherent to rare pediatric brain tumors. Subgroup analysis comparing PET scanner hardware generations was unfeasible because most patients were scanned on the newer Discovery MI system, though ratio-based metrics buffered scanner discrepancies. Furthermore, the neural network was fine-tuned on expert-reviewed semi-automatic segmentations rather than manual contours, reflecting the difficulty of generating manual three-dimensional ground truths in pediatric clinical environments. Dynamic PET findings also remain exploratory due to the thirty-two-patient sample. To support clinical research and external reproduction, the authors released the complete workflow within an open-source graphical user interface named GliAAns-UI, designed in Python 3.11 and hosted on GitHub. Future efforts will focus on expanding datasets through multicentric collaborations to validate dynamic parameters and evaluate longitudinal therapy monitoring.

Subject of Research: Deep learning-based automated tumour segmentation and PET parameter extraction in paediatric gliomas

Article Title: U-Net-based transfer learning for automated tumour segmentation enabling fully automated [18F]F-DOPA PET analysis in paediatric gliomas

Article References: Mureddu, M., Trò, R., Garau, F. G., Trebino, N., Bianconi, A., Rossi, A., Ramaglia, A., Verrico, A., Milanaccio, C., Morana, G., Iacozzi, M., Fiz, F., Piccardo, A., & Fato, M. M. (2026). U-Net-based transfer learning for automated tumour segmentation enabling fully automated [18F]F-DOPA PET analysis in paediatric gliomas. Brain Informatics, 13(1), Article 10. https://doi.org/10.1186/s40708-026-00296-z

Image Credits: AI Generated

DOI: 10.1186/s40708-026-00296-z

Keywords: Paediatric gliomas, [18F]F-DOPA PET, Transfer learning, U-Net architecture, Tumour segmentation, FLAIR MRI, Tumour-to-striatum ratio, Tumour-to-background ratio, Dynamic PET imaging, BraTS dataset, Overall survival, Progression-free survival, Histological tumour grade, Time-to-peak, Dynamic slope ratio, GliAAns-UI, Neuro-oncology, Medical image processing

News Source: Harold Sullivan. (November 12, 2026). Transfer Learning Framework Automates Paediatric Glioma Segmentation And PET Biomarker Analysis. Scienmag.

Tags: BraTS datasetDynamic PET imagingDynamic slope ratioFLAIR MRIGliAAns-UIHistological tumour grademedical image processingNeuro-oncologyoverall survivalPaediatric gliomasprogression-free survivalTime-to-peakTransfer LearningTumour segmentationtumour-to-background ratioTumour-to-striatum ratioU-Net++ architecture[18F]F-DOPA PET
Share12Tweet7Share2ShareShareShare1

Related Posts

Ketogenic Diet Improves Renal Fanconi Syndrome in Single Pediatric Cystinosis Case

Ketogenic Diet Improves Renal Fanconi Syndrome in Single Pediatric Cystinosis Case

November 17, 2026
Self-Supervised Framework Enhances Anomaly Detection Accuracy Across Dynamic Internet of Things Networks

Self-Supervised Framework Enhances Anomaly Detection Accuracy Across Dynamic Internet of Things Networks

November 17, 2026

Most Patients Achieve Pouch Preservation Using Endoscopic or Revisional Approaches, Narrative Review Finds

November 17, 2026

AI Model Improves Geological Pattern Discovery Through Causal Chain Analysis

November 17, 2026

POPULAR NEWS

  • Vitamin A Deficiency Persists in Congolese Children Despite Long-Term Supplementation Efforts

    Vitamin A Deficiency Persists in Congolese Children Despite Long-Term Supplementation Efforts

    29 shares
    Share 12 Tweet 7
  • Veterinarians Identify Twelve Defences to Improve Medication Safety in Sows

    29 shares
    Share 12 Tweet 7
  • Ascending Aorta Shows Marked Elongation Before Dissection Despite Only Moderate Dilatation

    29 shares
    Share 12 Tweet 7
  • Rural Australian Hospital Study Links Insulin and Polypharmacy to Medication Readmissions

    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

Physicians Prioritize Survival Benefits in Treating Advanced Non–Small Cell Lung Cancer

Ketogenic Diet Improves Renal Fanconi Syndrome in Single Pediatric Cystinosis Case

Self-Supervised Framework Enhances Anomaly Detection Accuracy Across Dynamic Internet of Things Networks

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.