Deep Learning-Assisted Quantitative Susceptibility Mapping as a Tool for Grading and Molecular Subtyping of Gliomas

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Neuro-oncology

Background:

• Accurate glioma grading and molecular subtyping are crucial for treatment planning and prognosis.
• Conventional MRI techniques have limitations in non-invasively assessing these tumor characteristics.
• Quantitative Susceptibility Mapping (QSM) offers potential for characterizing tissue magnetic properties.

Purpose of the Study:

• To evaluate the efficacy of deep learning (DL)-assisted QSM in glioma grading.
• To assess the utility of DL-assisted QSM in molecular subtyping of gliomas, specifically predicting IDH1 and ATRX status.
• To compare the performance of DL-assisted QSM with conventional MRI sequences (T2 FLAIR, T1WI+C).

Main Methods:

• Retrospective analysis of 42 patients with gliomas undergoing 3.0T MRI, including T2 FLAIR, T1WI+C, and QSM.
• Tumor segmentation performed manually; an inception convolutional neural network (CNN) used for feature extraction from MRI slices.
• Fivefold cross-validation employed for training and evaluation of model performance using accuracy and AUC.

Main Results:

• Single-modal QSM with DL demonstrated superior performance over T2 FLAIR and T1WI+C in differentiating glioblastomas from other grades and predicting IDH1/ATRX status.
• Combining QSM, T2 FLAIR, and T1WI+C with DL achieved the highest accuracy in glioma grading (0.89) and predicting IDH1 mutation (0.89) and ATRX loss (0.71).
• DL-assisted QSM significantly outperformed individual conventional MRI modalities in key diagnostic tasks.

Conclusions:

• Deep learning-assisted QSM is a valuable tool for non-invasive glioma grading and molecular subtyping.
• This AI-driven approach complements conventional MRI, offering improved diagnostic accuracy for critical glioma characteristics.
• DL-assisted QSM holds promise as a novel molecular imaging technique in neuro-oncology.