Breast mass segmentation in ultrasound with selective kernel U-Net convolutional neural network

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate breast mass segmentation in ultrasound (US) images is challenging due to variations in size and image characteristics.
• Deep learning methods offer potential for automating this complex task.

Purpose of the Study:

• To develop and evaluate a novel deep learning method for improved breast mass segmentation in ultrasound images.
• To address segmentation difficulties caused by variations in breast mass size and image properties.

Main Methods:

• Development of a selective kernel (SK) U-Net convolutional neural network.
• Utilizing an attention mechanism in SKs to adjust receptive fields and fuse feature maps from dilated and conventional convolutions.
• Training and evaluation on a large dataset of 882 breast masses and testing on three external datasets (893 US images).

Main Results:

• The SK-U-Net achieved a mean Dice score of 0.826 on the test set, outperforming the regular U-Net (0.778).
• Performance on external datasets ranged from 0.646 to 0.780.
• Fine-tuning with multi-center data improved performance by approximately 6%.
• A strong correlation (Spearman's rank coefficient of 0.7) was found between dilated convolutions and breast mass size.

Conclusions:

• Deep learning, specifically the SK-U-Net, demonstrates significant utility for breast mass segmentation in ultrasound imaging.
• The proposed method offers enhanced accuracy and robustness across different datasets.
• The study highlights the effectiveness of combining dilated and conventional convolutions for improved segmentation.