Brief Review — Segmentation and Recognition of Breast Ultrasound Images Based on an Expanded U-Net

Expanded U-Net, Expanding U-Net from 2-Class to 3-Class Prediction

Ultrasound image and label of breast tumor. Picture (a) is a benign tumor, and picture (b) is a malignant tumor.

Segmentation and Recognition of Breast Ultrasound Images Based on an Expanded U-Net,
Expanded U-Net, by Beijing Institute of Technology, Chinese Academy of Sciences, and Chengcheng County Hospital,
2021 J PLoS ONE (Sik-Ho Tsang @ Medium)
Medical Imaging, Medical Image Analysis, Image Segmentation, U-Net

• An image segmentation system for breast ultrasound images is proposed using an expanded U-Net.

Outline

1. Expanded U-Net
2. Results

1. Expanded U-Net

Expanded U-Net architecture

1.1. Architecture

• U-Net, with the above architecture, is used.
• In this paper, 7 Dropout layers were added to the original U-Net.
• The Batch Normalization technique is used after each convolution and deconvolution, and Leaky ReLU is used.
• In the last layer, softmax is selected as the activation function to map the 8-channel feature vector to the 3-channel grey-level output map.

1.2. Grey-Level Probability Label

Label of the expanded U-Net dataset.

• The ground truth is converted to a three-channel greyscale image with a grey level of 256 and then normalized to grey-level probability labels of 0, 0.5 and 1.
• “Expanded” means expanding binary cross-entropy function into ternary cross-entropy function:

2. Results

U-Net quantitative analysis results by expanded training and general training.

• The Dice coefficient of the expanded U-Net is 7.6 larger than that of the general U-Net.
• The IOU coefficient of the expanded U-Net is 11.0 larger than that of the general U-Net.

Comparison of test results in breast ultrasound images.

• Comparing the third and fourth columns, it can be seen that the output map of the expanded U-Net retains the texture details and edge features of the breast tumor.

Although the title is about segmentation and recognition, it is a 3-class segmentation problem.

Reference

[2021 J PLoS ONE] [Expanded U-Net]
Segmentation and Recognition of Breast Ultrasound Images Based on an Expanded U-Net

4.2. Biomedical Image Segmentation

2015 … 2021 [Expanded U-Net]

==== My Other Previous Paper Readings ====