Modifying the U-Net's Encoder-Decoder Architecture for Segmentation of Tumors in Breast Ultrasound Images.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer-Aided Diagnosis

Background:

• Accurate segmentation of breast ultrasound images is crucial for early cancer detection.
• Challenges in segmentation include speckle noise, low signal-to-noise ratio, and intensity heterogeneity.
• Existing methods struggle with the complexities of ultrasound image artifacts.

Purpose of the Study:

• To develop an improved method for accurate and effective breast ultrasound image segmentation.
• To enhance the early diagnosis of breast cancer through precise medical image analysis.
• To introduce a novel neural network architecture for superior lesion segmentation.

Main Methods:

• Proposed a neural network (NN) based on U-Net and an encoder-decoder architecture, named CRes-U-Net.
• Combined U-Net with Res-Net and MultiResUNet, incorporating a novel 'Co-Block' to preserve low-level and high-level features.
• Evaluated the network on the Breast Ultrasound Images (BUSI) Dataset comprising 780 images (normal, benign, malignant).

Main Results:

• The CRes-U-Net network achieved high performance metrics on the BUSI dataset.
• Achieved 82.88% Dice Similarity Coefficient (DSC), 77.5% Intersection over Union (IoU), 90.3% Area Under Curve (AUC), and 98.4% global accuracy (ACC).
• Demonstrated superior segmentation accuracy compared to other state-of-the-art deep learning methods.

Conclusions:

• The proposed CRes-U-Net effectively segments breast lesions in ultrasound images.
• This method offers improved accuracy and effectiveness for computer-aided diagnosis of breast cancer.
• The novel architecture successfully addresses challenges posed by ultrasound image artifacts and noise.