Automatic Brain Tumor Segmentation Based on Cascaded Convolutional Neural Networks With Uncertainty Estimation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Accurate brain tumor segmentation is crucial for treatment planning.
• Existing convolutional neural networks (CNNs) face challenges balancing 3D context with memory consumption.
• Current methods often lack uncertainty quantification for segmentation results.

Purpose of the Study:

• To develop an efficient and accurate method for segmenting brain tumors and their subregions from multi-modal MRI.
• To introduce a 2.5D CNN approach as a balance between memory usage and contextual information.
• To incorporate test-time augmentation for improved accuracy and uncertainty estimation.

Main Methods:

• A cascaded framework of 2.5D CNNs was proposed for hierarchical brain tumor segmentation.
• Test-time augmentation was employed to enhance segmentation accuracy and generate uncertainty maps.
• The method was evaluated on the BraTS 2017 and BraTS 2018 datasets.

Main Results:

• The proposed cascaded 2.5D CNN framework achieved a second-rank performance in the BraTS 2017 challenge.
• Test-time augmentation demonstrated significant improvements in brain tumor segmentation accuracy.
• Generated uncertainty information effectively highlighted potential segmentation errors.

Conclusions:

• The cascaded 2.5D CNN approach offers an effective solution for brain tumor segmentation, balancing performance and computational resources.
• Test-time augmentation enhances segmentation accuracy and provides crucial uncertainty metrics.
• The uncertainty information can aid in refining segmentation and improving clinical decision-making.