AUTOMATIC BRAIN ORGAN SEGMENTATION WITH 3D FULLY CONVOLUTIONAL NEURAL NETWORK FOR RADIATION THERAPY TREATMENT...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiation Oncology

Background:

• 3D organ contouring is crucial for radiation therapy planning, impacting dose estimation and risk reduction.
• Manual contouring is labor-intensive and suffers from inter-observer variability, affecting treatment outcomes.
• Significant variations in organ sizes pose a challenge for accurate segmentation.

Purpose of the Study:

• To introduce BrainSegNet, a novel 3D fully convolutional neural network (FCNN) for automated brain organ segmentation.
• To address the challenge of large organ size variability using a multi-resolution approach and weighted loss function.
• To evaluate BrainSegNet's performance against existing methods for brain organ segmentation.

Main Methods:

• Developed BrainSegNet, a 3D FCNN incorporating multiple resolution paths.
• Implemented a weighted loss function to handle diverse organ sizes.
• Validated the model using 46 Brain CT image volumes with expert-annotated contours.

Main Results:

• BrainSegNet demonstrated superior performance in segmenting both small/thin organs (chiasm, optic nerves, cochlea) and large organs.
• Achieved higher accuracy compared to LiviaNet and V-Net.
• Reduced manual contouring time from approximately one hour to under two minutes per volume.

Conclusions:

• BrainSegNet offers an effective and efficient solution for automated brain organ segmentation in radiation therapy.
• The method shows high potential for improving the overall efficiency of radiation therapy workflows.
• Accurate segmentation of organs at risk is vital for personalized and safe radiation treatments.