A research team led by Prof. HU Zhanli from the Shenzhen Institute of Advanced Technology (SIAT) of the Chinese Academy of Sciences, has proposed an automated whole-brain segmentation approach based on the cross-attention mechanism designed for positron emission tomography/magnetic resonance imaging (PET/MRI) bimodal images.

The study was published in IEEE Transactions on Radiation and Plasma Medical Sciences on Jun 13.

Brain is a crucial organ in the human body, with its various regions controlling a multitude of physiological functions. Abnormalities within these brain regions can have systemic effects, potentially impacting the function of other organs and systems.

The integrated PET/MRI system has emerged as a pivotal and promising tool for detecting brain diseases, enabling the simultaneous acquisition of both biological and anatomical information, thereby providing comprehensive diagnostic and research insights. Whole-brain segmentation plays a significant role in quantifying and visualizing anatomical structures, planning surgical procedures, and performing image-guided interventions.

However, manual segmentation is labor-intensive and depends on the expertise of experienced radiologists. When dealing with large datasets, this method becomes significantly costly, time-consuming, and impractical due to variations in experience and knowledge among the experts involved.

In this study, the researchers used PET/MRI dual-modality brain images obtained from 120 patients. They developed a novel 3D network with cross-attention module, enabling them to capture the correlation between dual-modality features and thereby improve the segmentation accuracy.

With the proposed approach, superior evaluation metrics (85.35% Dice, 77.22% Jaccard, 88.86% Precision and 84.81% Recall) were achieved. Moreover, correlation and consistency analyses based on different brain regions demonstrated that the proposed approach maintains a consistent data distribution shape with the ground truth and enables the preservation of correlation relationships across the whole brain.

This study sheds light on the possibility of applying automatic whole-brain segmentation incorporating functional and anatomical information, and shows potential for promoting the development of personalized treatment and precision medicine for related brain diseases.

Correlation and consistency analyses results on PET SUV distributions within different brain regions (Image by SIAT)

Media Contact: LU Qun

Email: qun.lu@siat.ac.cn

Download the attachment： 

   Accurate Whole-Brain Segmentation for Bimodal PET/MR Images Via a Cross-Attention Mechanism