Brief Review — An Efficient Solution for Breast Tumor Segmentation and Classification in Ultrasound Images Using Deep Adversarial Learning

Conditional GAN (cGAN) + Atrous Convolution (AC) + Channel Attention with Weighting Block (CAW)

4 min readNov 8, 2022

An Efficient Solution for Breast Tumor Segmentation and Classification in Ultrasound Images Using Deep Adversarial Learning
cGAN+AC+CAW, by Universitat Rovira i Virgili, and Bioinformatics Institute,
2019 arXiv v1 (Sik-Ho Tsang @ Medium)
Medical Image Analysis, Image Segmentation, Image Classification

An efficient solution for tumor segmentation and classification in breast ultrasound (BUS) images. An atrous convolution layer is added to the conditional generative adversarial network (cGAN) segmentation model. A channel-wise weighting block is also added.
SSIM and L1-norm loss with the typical adversarial loss are proposed as a loss function.

Outline

cGAN+AC+CAW for Segmentation
Random Forest for Classification
Results

1. cGAN+AC+CAW for Segmentation

1.1. Generator G

The generator network incorporates an encoder section, made of seven convolutional layers (En1 to En7), and a decoder section, made of seven deconvolutional layers (Dn1 to Dn7).
An atrous convolution block is inserted between En3 and En4. 1, 6 and 9 dilation rates with kernel size 3×3 and a stride of 2 are used.
in addition to a channel attention with channel weighting (CAW) block between En7 and Dn1.
The CAW block is an aggregation of a channel attention module (DAN) with channel weighting block (SENet), which increases the representational power of the highest level features of the generator network.

1.2. Discriminator D

It is a sequence of convolutional layers.
The input of the discriminator is the concatenation of the BUS image and a binary mask marking the tumor area.
The output of the discriminator is a 10×10 matrix having values varying from 0.0 (completely fake) to 1.0 (real).

1.3. Loss Function

The loss function of the generator G comprises three terms: adversarial loss (binary cross entropy loss), L1-norm to boost the learning process, and SSIM loss to improve the shape of the boundaries of segmented masks:

where z is a random variable.
The loss function of the discriminator D is:

2. Random Forest for Classification

Each BUS image is fed into the trained generative network to obtain the boundary of the tumor, and then 13 statistical features from that boundary are computed: fractal dimension, lacunarity, convex hull, convexity, circularity, area, perimeter, centroid, minor and major axis length, smoothness, Hu moments (6) and central moments (order 3 and below).
Exhaustive Feature Selection (EFS) algorithm is used to select the best set of features. The EFS algorithm indicates that the fractal dimension, lacunarity, convex hull, and centroid are the 4 optimal features.
The selected features are fed into a Random Forest classifier, which is later trained to discriminate between benign and malignant tumors.

3. Results

3.1. Segmentation

This dataset contains 150 malignant and 100 benign tumors contained in BUS images. To train our model, we randomly divided the dataset into the training set (70%), a validation set (10%) and testing set (20%).

The model (cGAN+AC+CAW) outperforms the rest in all metrics. It achieves Dice and IoU scores of 93.76% and 88.82%, respectively.

**Boxplots of IoU and Dice metrics of the proposed model and** **FCN**, **SegNet**, **ERFNet** **and** **U-Net**.

The proposed model is in the range 88% to 94% for Dice coefficient and 80% to 89% for IoU, while other deep segmentation methods, FCN, SegNet, ERFNet and U-Net show a wider range of values.

**Segmentation results on four samples of the BUS dataset.**

SegNet and ERFNet yield the worst results since there are large false negative areas (in red), as well as some false positive areas (in green).