Ultrasound Image Segmentation: A Deeply Supervised Network With Attention to Boundaries

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Manual segmentation of ultrasound images is subjective and requires expertise.
• Automated segmentation methods are needed to overcome limitations of manual approaches.

Purpose of the Study:

• To develop a fully convolutional neural network (FCNN) with attentional deep supervision for accurate automatic segmentation of ultrasound images.
• To improve the segmentation of anatomical structures in ultrasound, addressing challenges like broken boundaries.

Main Methods:

• A fully convolutional neural network (FCNN) with sub-problem specific deep supervision was developed.
• Attention mechanisms guided fine resolution layers for boundary learning, while coarse layers focused on region discrimination.
• Customized loss down-weighting and a trainable fusion layer were introduced.

Main Results:

• The proposed network achieved superior performance in blood vessel segmentation compared to existing methods (F1 score 0.83, mIoU 0.83, Dice 0.79).
• Achieved a Dice index of 0.91 for lumen segmentation on the MICCAI 2011 IVUS dataset, closely matching the reference value.
• Demonstrated significant improvements in lesion segmentation, comparable to vessel segmentation results.

Conclusions:

• Sub-problem specific deep supervision significantly improves overall segmentation accuracy in ultrasound images.
• The proposed method effectively handles challenges like broken boundaries common in ultrasound imaging.