A multiple-channel and atrous convolution network for ultrasound image segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Ultrasound image segmentation is critical for medical diagnosis but is hindered by low signal-to-noise ratios and poor image quality.
• Existing convolutional neural network (CNN) approaches exhibit limited generalization capabilities for ultrasound image segmentation.

Purpose of the Study:

• To develop an end-to-end, multiple-channel and atrous CNN, termed MA-Net, for improved semantic information extraction in ultrasound image segmentation.
• To enhance the generalization ability of CNNs in segmenting challenging ultrasound images.

Main Methods:

• Developed MA-Net, an encoder-decoder architecture incorporating multiple-channel convolution, large kernels, atrous convolution, pyramid pooling, and a residual skip pathway.
• Employed multi-task learning and extensive data augmentation techniques to optimize segmentation performance.
• Evaluated segmentation using Dice score, precision, recall, Hausdorff distance (HD), average symmetric surface distance (ASD), and root mean square symmetric surface distance (RMSD), with Friedman test for statistical analysis.

Main Results:

• MA-Net achieved superior segmentation performance across brachia plexus, fetal head, and lymph node datasets, outperforming U-Net, U-Net++, M-Net, and Dilated U-Net.
• Demonstrated significant improvements in Dice score (up to 36.03%), precision, recall, HD, ASD, and RMSD compared to existing methods.
• Validated MA-Net's generalization on lower grade brain glioma MRI and lung CT images, achieving the highest mean rank in the Friedman test.

Conclusions:

• The proposed MA-Net provides accurate ultrasound image segmentation with high generalization capabilities.
• MA-Net represents a valuable tool for improving diagnostic accuracy in ultrasound imaging applications.