Deep Learning Methodology for Differentiating Glioma Recurrence From Radiation Necrosis Using Multimodal Magnetic...

Area of Science:

• Neuro-oncology
• Radiology
• Artificial Intelligence in Medicine

Background:

• Distinguishing glioma recurrence from radiation necrosis (pseudoprogression) is critical for patient management.
• Accurate radiological assessment is essential for treatment planning and prognosis.

Purpose of the Study:

• To develop and validate a deep learning (DL) methodology for automated differentiation of tumor recurrence from radiation necrosis.
• To assess the performance of the DL model using routine magnetic resonance imaging (MRI) scans.

Main Methods:

• A retrospective study analyzed 146 glioma patients with suspected recurrent lesions on follow-up MRI.
• A light-weighted deep neural network (ERN-Net) was trained on T1, T2, and contrast-enhanced T1 MRI sequences.
• Model performance was evaluated using sensitivity, specificity, accuracy, and AUC, and compared to neurosurgeons.

Main Results:

• Multimodal MRI-based DL models outperformed single-modal models.
• ERN-Net achieved high AUC values (0.915 image-wise, 0.958 subject-wise).
• DL models demonstrated superior sensitivity, specificity, and accuracy compared to experienced neurosurgeons.

Conclusions:

• Deep learning provides a valuable computational tool for differentiating recurrent gliomas from radiation necrosis.
• The ERN-Net model shows excellent performance and high clinical applicability on routine MRI scans.
• This AI-driven approach can aid in the accurate diagnosis and management of glioma patients.