نهمین کنگره نقشه برداری مغز

نمایش مقاله

Title:

Early Diagnosis of Diffuse Midline Glioma Using Deep Learning: A Systematic Review

Subject:

Personalized Interventions and Diagnosis of Complex Neurological Disorders

Presenter Author:
Name:	PARMIDA SERAJ

Author:
Name 1:	PARMIDA SERAJ - Corresponding - Presenter
Affiliation 1:	Medical student, Department of Medicine, Tehran Medical Branch, Islamic Azad University, Tehran, Iran

Name 2:	Sadra Hoseini
Affiliation 2:	Medical student, Department of Medicine, Tehran Medical Branch, Islamic Azad University, Tehran, Iran

Name 3:	Mohammad souri
Affiliation 3:	Student Research Committee, Lorestan University of Medical Sciences, Khorramabad, Iran.

Name 4:
Affiliation 4:

Name 5:
Affiliation 5:

Name 6:
Affiliation 6:

Name 7:
Affiliation 7:

Name 8:
Affiliation 8:

Name 9:
Affiliation 9:

Name 10:
Affiliation 10:

Abstract

Original Research

introduction :

Diffuse midline gliomas (DMG), including diffuse intrinsic pontine glioma (DIPG), are rare, aggressive pediatric brain tumors with dismal survival. Early and accurate diagnosis is critical to guide therapy and trial enrollment. While MRI is the cornerstone of initial assessment, subtle phenotypes and overlap with other midline lesions complicate timely recognition and molecular characterization. Artificial intelligence (AI) and deep learning (DL) can mine quantitative imaging features (radiomics) from MRI and diffusion sequences and, when available, digital histopathology, potentially enabling non-invasive classification and risk stratification. In this review, “early detection” is operationalized as AI-assisted identification, differential diagnosis, or molecular prediction at or before the initial diagnostic work-up, prior to definitive treatment decisions. We summarize current evidence on AI/DL for DMG with attention to diagnostic performance, prognostication, and methodological quality.

Methods:

We searched PubMed and Embase for English-language studies published from January 2015 to July 2025 combining terms for DMG/DIPG/brainstem glioma, imaging modalities (MRI, diffusion/DTI, radiomics, histopathology), and AI/ML techniques (deep learning, convolutional neural networks, transfer learning, support vector machines, ensemble methods) alongside early detection or prognostication. Preprints and reference lists were additionally screened. After deduplication, 69 unique records were screened by two independent reviewers, with disagreements resolved by discussion or a third reviewer. Inclusion required original research applying AI/ML to pre-diagnostic detection, classification, molecular prediction, segmentation, or prognosis in DMG using MRI, DTI, or histopathology; reviews, editorials, and case reports were excluded. Twelve studies met criteria and were synthesized qualitatively. Given heterogeneity in populations, imaging protocols, feature engineering, and reported metrics, meta-analysis was not attempted; where available, we summarize discrimination, calibration, and validation strategies.

Results:

Across the 12 studies, AI models addressed three main tasks: (i) differential diagnosis among midline lesions and automated segmentation; (ii) non-invasive prediction of H3K27-altered status; and (iii) prognostication, most commonly overall survival (OS). Reported performance for classification was generally favorable, and several studies showed that DL/ML could distinguish DMG from other midline tumors on conventional MRI radiomics. Predictive models for H3K27 alteration achieved promising accuracy, suggesting a potential role in triaging and prioritizing biopsy and molecular testing rather than replacing tissue diagnosis. Prognostic modeling was less consistent; one study reported 83.3% accuracy for early OS prediction. Nonetheless, evidence quality is limited: most cohorts were single-center with small sample sizes (approximately 30–70 patients), reporting was methodologically heterogeneous, and key metrics (e.g., AUC, sensitivity/specificity, calibration, and external validation) were inconsistently provided. Few studies detailed steps to mitigate overfitting or data leakage, and variability in MRI acquisition and feature extraction hindered comparability. Collectively, these gaps constrain generalizability and preclude pooled effect estimates.

Conclusion:

DL and radiomics are promising adjuncts for early diagnostic triage in DMG, but robust, multi-center, externally validated, prospectively designed studies with standardized reporting and reproducibility are required before clinical adoption.

Keywords:

Diffuse midline glioma (DMG); Deep learning; Artificial intelligence; MRI; Diffusion tensor imaging