Related Experiment Video
Updated: Feb 28, 2026
10:25
Brain Infarct Segmentation and Registration on MRI or CT for Lesion-symptom Mapping
Published on: September 25, 2019
49.5K
IDH Mutation Assessment in Gliomas from Anatomical MRI Using Deep Learning: A Comparative Analysis of Centralized and
Abdullah Bas1, Esin Ozturk-Isik1,2
1Institute of Biomedical Engineering, Bogazici University, Istanbul 34684, Türkiye.
Diagnostics (Basel, Switzerland)
|February 27, 2026
Summary
Deep learning models can accurately detect isocitrate dehydrogenase (IDH) mutations in diffuse gliomas using standard MRI scans. Context-preserving image preprocessing significantly improved model performance, offering a non-invasive diagnostic alternative.
Area of Science:
- Artificial Intelligence in Medical Imaging
- Neuro-oncology
- Radiomics and Computational Pathology
Background:
- Isocitrate dehydrogenase (IDH) mutation status is critical for diffuse glioma prognosis but requires invasive tissue sampling.
- Non-invasive preoperative identification of IDH mutations via routine anatomical MRI could guide clinical decisions.
- This study investigates deep learning for IDH mutation detection using MRI and evaluates preprocessing and training strategies.
Purpose of the Study:
- To evaluate deep learning models for non-invasive IDH mutation detection in diffuse gliomas using anatomical MRI.
- To compare the impact of tumor-focused image preprocessing techniques (NSF vs. GBSF) on model performance.
- To assess the performance of centralized learning (CL) versus federated learning (FL) with different aggregation strategies (FA vs. FTM).
Main Methods:
- A deep learning classifier (2D U-Net encoder) was developed using anatomical MRI (T1c, T2, FLAIR) from 501 diffuse glioma patients.
- Two preprocessing methods, Naïve Soft Filtering (NSF) and Gradient-Based Soft Filtering (GBSF), were applied to MRI data.
- Model performance was compared between CL and FL (using FA and FTM) based on accuracy, F1 score, and ROC-AUC.
Main Results:
- The centralized learning (CL) model with Naïve Soft Filtering (NSF) achieved the highest performance (accuracy=0.949, F1=0.951, ROC-AUC=0.971).
- NSF consistently outperformed GBSF across all evaluated training schemes.
- Federated learning (FL) showed decreased performance compared to CL, with Federated Averaging (FA) outperforming Federated Trimmed Mean (FTM).
Conclusions:
- Deep learning models can accurately classify IDH mutation status from routine anatomical MRI, offering a non-invasive diagnostic approach.
- Context-preserving preprocessing with NSF significantly enhances model performance regardless of the training scheme.
- Federated learning offers a privacy-preserving alternative but requires careful selection of aggregation strategies to minimize performance loss.
