Brain tumor segmentation using holistically nested neural networks in MRI images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Gliomas are aggressive brain tumors requiring precise imaging for diagnosis and treatment.
• Magnetic Resonance Imaging (MRI) is crucial for glioma management but segmentation remains challenging.
• Accurate tumor delineation is vital for effective radiation therapy planning.

Purpose of the Study:

• To present a novel automatic brain tumor segmentation method using holistically nested neural networks (HNN).
• To improve the accuracy and efficiency of brain tumor segmentation in MRI images.
• To develop a practical tool for clinical glioma diagnosis and radiotherapy planning.

Main Methods:

• Applied N4ITK for bias field correction and developed a landmark-based intensity normalization.
• Utilized holistically nested neural networks (HNN), an extension of CNNs, for multiscale tumor feature learning.
• Trained HNN on BRATS 2013 datasets and clinical data, followed by optimum thresholding for segmentation.

Main Results:

• Achieved a Dice Similarity Coefficient (DSC) of 0.78 and sensitivity of 0.81 on BRATS 2013 data.
• Obtained DSC of 0.83 and sensitivity of 0.85 on local clinical datasets.
• Demonstrated superior performance and efficiency compared to the Fully Convolutional Network (FCN) method.

Conclusions:

• Developed an effective HNN-based method for brain tumor segmentation in MRI.
• The method's accuracy and efficiency make it practical for clinical applications.
• This technique holds potential for enhancing brain tumor diagnostic analysis and radiation therapy planning.