A low resource 3D U-Net based deep learning model for medical image analysis.

Area of Science:

• Artificial Intelligence (AI)
• Deep Learning (DL)
• Medical Image Analysis
• Radiology

Background:

• Deep learning shows promise for enhancing clinical decision support systems in radiology.
• Accurate brain tumor segmentation is crucial for diagnosis, treatment planning, and survival prediction.
• Gliomas present challenges due to irregular shapes and ambiguous boundaries, often requiring multimodal imaging analysis.

Purpose of the Study:

• To present a fully automatic deep learning method for brain tumor segmentation.
• To segment tumors in multimodal, multi-contrast magnetic resonance imaging (MRI) scans.
• To develop a computationally efficient model suitable for diverse healthcare environments.

Main Methods:

• A lightweight UNet architecture was employed, featuring a multimodal Convolutional Neural Network (CNN) encoder-decoder model.
• The method was evaluated using the Brain Tumor Segmentation (BraTS) Challenge 2018 dataset.
• The approach requires no data augmentation and minimal computational resources.

Main Results:

• The proposed lightweight UNet model achieved improved performance compared to previous challenge models.
• The model demonstrated effectiveness without relying on extensive data augmentation or heavy computational infrastructure.
• The segmentation accuracy was enhanced, providing a reliable computer-aided diagnosis tool.

Conclusions:

• The developed deep learning model offers an effective solution for automatic brain tumor segmentation.
• Its efficiency and reduced resource requirements make it highly suitable for remote and low-resource healthcare settings.
• This AI-driven approach can significantly aid radiologists in timely and accurate clinical diagnosis.