One-pass Multi-task Networks with Cross-task Guided Attention for Brain Tumor Segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computational Neuroscience

Background:

• Class imbalance is a significant challenge in medical image segmentation.
• Model cascade (MC) strategies improve segmentation but introduce complexity and ignore inter-model correlations.

Purpose of the Study:

• To propose a light-weight, one-pass deep model (OM-Net) for brain tumor segmentation that overcomes the limitations of MC strategies.
• To enhance segmentation accuracy and efficiency by addressing class imbalance more effectively.

Main Methods:

• Developed OM-Net, a multi-task network with shared and task-specific parameters for joint feature learning.
• Implemented online training data transfer and curriculum learning strategies to optimize OM-Net.
• Introduced a cross-task guided attention (CGA) module for adaptive feature recalibration.
• Utilized a post-processing method to refine segmentation results.

Main Results:

• OM-Net achieved state-of-the-art performance on the BraTS 2015 and BraTS 2017 datasets.
• The proposed methods demonstrated superior effectiveness in brain tumor segmentation.
• Achieved joint third place in the BraTS 2018 challenge.

Conclusions:

• OM-Net offers a more efficient and effective solution for brain tumor segmentation compared to existing methods.
• The integration of multi-task learning, attention mechanisms, and optimized training strategies significantly improves segmentation performance.
• The developed approach provides a robust framework for tackling class imbalance in medical image analysis.