Radiomics-Driven Hybrid Deep Learning for MRI-Based Prediction of Glioma Grade and 1p/19q Codeletion.

Area of Science:

• Neuro-oncology
• Radiomics
• Artificial Intelligence in Medicine

Background:

• Accurate preoperative glioma grading and molecular profiling are crucial for personalized treatment.
• 1p/19q codeletion status is a key prognostic marker in low-grade gliomas (LGGs).
• Current assessment methods are invasive, necessitating non-invasive alternatives.

Purpose of the Study:

• To develop and validate a non-invasive radiomics framework for glioma grading.
• To predict the 1p/19q codeletion status using quantitative MRI features and machine learning.
• To compare the performance of traditional ML and deep learning models for these predictions.

Main Methods:

• Extracted high-dimensional radiomic features (geometry, intensity, texture) from preoperative MRI.
• Employed feature selection for normalization and optimization.
• Compared traditional ML classifiers with deep learning models, including CNNs, RNNs, and a hybrid CNN-LSTM model.
• Validated models using five-fold cross-validation and an independent test set.

Main Results:

• The hybrid CNN-LSTM model achieved the highest performance with 88.1% accuracy and 0.93 AUC.
• This hybrid deep learning model outperformed conventional ML and single deep learning architectures.
• Explainability analysis highlighted tumor heterogeneity and morphology features as most impactful.

Conclusions:

• Radiomic features combined with hybrid deep learning models can non-invasively predict glioma grade and 1p/19q codeletion status.
• This computational model shows potential as a supplementary tool in precision neuro-oncology.
• Non-invasive prediction can aid in tailoring treatment strategies for glioma patients.