QuickNAT: A fully convolutional network for quick and accurate segmentation of neuroanatomy

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Image Analysis

Background:

• Whole brain segmentation of structural magnetic resonance imaging (MRI) is essential for morphological analysis but is computationally intensive.
• The computational demands of segmentation can delay the availability of critical imaging biomarkers post-scan acquisition.

Purpose of the Study:

• To introduce QuickNAT, a highly efficient neural network for rapid whole brain segmentation of MRI scans.
• To develop a training strategy that leverages large neuroimaging datasets with limited manual annotations.

Main Methods:

• QuickNAT utilizes a fully convolutional, densely connected neural network architecture.
• A novel training approach involves pre-training on auxiliary labels from existing software, followed by fine-tuning on manual labels to correct errors.
• This method enables effective training using large repositories without requiring extensive manual annotations.

Main Results:

• QuickNAT achieves whole brain segmentation in approximately 20 seconds.
• Evaluations across eight diverse datasets demonstrate superior segmentation accuracy and reliability compared to state-of-the-art methods.
• The network provides an orders-of-magnitude speed improvement over existing techniques.

Conclusions:

• QuickNAT offers a computationally efficient and accurate solution for whole brain MRI segmentation.
• The rapid processing facilitates the analysis of large-scale neuroimaging data and accelerates the clinical translation of imaging biomarkers.
• This technology supports faster clinical decision-making by providing near real-time image analysis.