STF-Net: sparsification transformer coding guided network for subcortical brain structure segmentation.

Area of Science:

• Medical Imaging
• Neuroscience
• Computer Vision

Background:

• Subcortical brain structure segmentation is crucial for neuroimaging diagnosis and computer-aided diagnosis.
• Manual segmentation is time-consuming, subjective, and limits clinical applications due to blurred boundaries and complex shapes.

Purpose of the Study:

• To propose an accurate and efficient method for subcortical brain structure segmentation.
• To address the limitations of manual segmentation in neuroimaging.

Main Methods:

• Introduced the sparsification transformer (STF) module incorporating a self-attention mechanism for global dependency extraction.
• Utilized a shallow network with convolutional operations for low-level detail compensation.
• Implemented a hybrid residual dilated convolution (HRDC) module for multi-scale contextual information and an octave convolution edge feature extraction (OCT) module for edge feature emphasis.
• Trained the network with a hybrid loss function.

Main Results:

• The proposed STF module demonstrated outstanding performance on the IBSR and MALC datasets.
• Achieved high accuracy in objective and subjective quality evaluations for brain structure segmentation.

Conclusions:

• The sparsification transformer (STF) module offers a promising solution for accurate and efficient subcortical brain structure segmentation.
• This method has the potential to significantly advance computer-aided diagnosis in neuroimaging.