CNN-based glioma detection in MRI: A deep learning approach

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Brain tumors, including high-grade gliomas (HGGs) and low-grade gliomas (LGGs), affect over a million people annually, posing significant diagnostic and treatment challenges.
• Accurate glioma segmentation in Magnetic Resonance Imaging (MRI) remains a critical hurdle in clinical practice, impacting patient prognosis.
• Convolutional neural networks (CNNs) show potential for improving segmentation accuracy and addressing the need for advanced diagnostic and therapeutic strategies.

Purpose of the Study:

• To develop an automated glioma segmentation algorithm utilizing CNNs for precise identification of tumor components in MRI.
• To achieve segmentation accuracy comparable to experienced radiologists and commercial tools, thereby enhancing diagnostic precision and quantification.

Main Methods:

• Analysis of 285 MRI scans from patients with HGGs and LGGs.
• Utilized T1-weighted (pre- and post-contrast) and T2-weighted (with and without FAIRE) sequences for segmentation.
• Employed a U-Net convolutional neural network for segmentation and computed DICE coefficients for tumor core, whole tumor, and tumor nucleus.

Main Results:

• The U-Net network achieved DICE coefficients of 0.7331 (tumor core with contrast), 0.8624 (entire tumor), and 0.7267 (tumor nucleus without contrast).
• Segmentation accuracy demonstrated comparability to professional radiologists and existing commercial segmentation tools.
• The CNN-based system effectively identified glioma components in MRI images.

Conclusions:

• A CNN-based automated segmentation system for gliomas was successfully developed, demonstrating high accuracy.
• The findings validate the capability of CNNs to improve the accuracy of brain tumor diagnoses.
• This advancement offers a promising direction for future research in medical imaging and diagnostics, potentially improving clinical workflows and patient care through precise, quantitative results.