Breast Ultrasound Images Segmentation Using Deep Neural Networks
Date
2024
Authors
Nguyen, Trong Khang
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Breast Ultrasound (US) imaging has emerged as an important diagnostic technique for detecting and characterizing breast tumors. Accurate segmentation of breast US images plays an essential role in enhancing the efficiency and precision of clinical assessments. This report explores the application of several well-known deep neural networks to the breast US image segmentation task. Specifically, we train and evaluate the following five models: SegNet, U Net, and DeepLab V3+ with three different bondnets (ResNet-18, ResNet-50, and Xception).
The presented results are based on two labeled datasets. One is Breast US Images (BUSI) dataset, which was used for training, validation, and testing. The other is Breast US Lesions (BUL) dataset, which was used exclusively for testing. Data augmentation was applied to increase the number and diversity of the data samples by randomly varying the contrast, brightness, and gamma of US images.
The performance of each model was evaluated based on Global Accuracy, Mean Accuracy, Mean Intersection-over-Union (IoU), Weighted IoU, Mean Boundary-F1 (BF) score, Average Dice score of Background, Average Dice score of Tumor, Mean Dice score of Tumor, and the model's cost. Overall, our results showed that Xception-based DeepLab V3+ and U-Net outperformed the other models under consideration when segmenting breast US images.
Description
Keywords
ultrasound, segmentation, neural networks, deep learning