A Multiscale Attentional Unet Model for Automatic Segmentation in Medical Ultrasound Images.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Deep Learning

Background:

• Ultrasonography is crucial for noninvasive clinical diagnosis.
• Accurate segmentation of regions of interest (ROIs) in ultrasound images is vital for computer-aided diagnosis (CAD).
• Segmentation of ROIs in low-contrast ultrasound images remains a significant challenge.

Purpose of the Study:

• To propose an efficient module, multiscale attentional convolution (MSAC), for improved medical ROI segmentation.
• To develop and evaluate MSAC-Unet, a novel deep learning architecture for ultrasound image segmentation.
• To assess the effectiveness of MSAC-Unet on thyroid nodules and brachial plexus nerves datasets.

Main Methods:

• Developed the multiscale attentional convolution (MSAC) module using cascaded convolutions and self-attention.
• Integrated MSAC into the Unet architecture, creating MSAC-Unet for encoder and decoder segmentation.
• Evaluated MSAC-Unet on thyroid nodule (TND-PUH3, DDTI) and brachial plexus nerve (NSD) ultrasound datasets.

Main Results:

• MSAC-Unet achieved high segmentation accuracy with Dice coefficients of 0.822 (TND-PUH3), 0.792 (DDTI), and 0.746 (NSD).
• The model demonstrated improved segmentation accuracy, yielding more reliable ROI edges and boundaries.
• Significantly reduced the number of erroneously segmented ROIs in ultrasound images.

Conclusions:

• MSAC-Unet effectively enhances ROI segmentation in challenging ultrasound images.
• The proposed MSAC module contributes to improved accuracy and reliability in medical image analysis.
• MSAC-Unet shows promise for advancing computer-aided diagnosis in ultrasonography.