BTS-GAN: Computer-aided segmentation system for breast tumor using MRI and conditional adversarial networks

• Published in: Engineering science and technology, an international journal Vol. 36; p. 101154
• Main Authors: Haq, Imran Ul, Ali, Haider, Wang, Hong Yu, Cui, Lei, Feng, Jun
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01-12-2022; Elsevier
• Subjects: Breast MRI; Breast tumor; Generative adversarial networks; Medical Image Segmentation; Parallel dilated convolution; Generative adversarial networks; Breast MRI; Parallel dilated convolution; Breast tumor; Medical Image Segmentation
• ISSN: 2215-0986; 2215-0986
• DOI: 10.1016/j.jestch.2022.101154

Breast tumor is one of the most prominent indicators for the diagnosis of breast cancer. The precise segmentation of tumors is crucial for enhancing the accuracy of breast cancer detection. A physician’s assessment of the MRI scan is time-consuming and require a lot of human effort and expertise. Furthermore, traditional medical segmentation approaches frequently need prior information or manual feature extraction, resulting in a subjective diagnosis. Therefore, the development of an automated image segmentation approach is essential for clinical applications. This work presents BTS-GAN, an automatic breast tumor segmentation process using conditional GAN (cGAN) in Magnetic Resonance Imaging (MRI) scans. First, we used an encoder-decoder deep network with skip connections between encoder and decoder for the generator to increase the localization efficiency. Second, we utilized a parallel dilated convolution (PDC) module to retain the features of various sizes of masses and to effectively extract information about the masses’ edges and interior texture. Third, an extra classification-related constraint is included to the loss function of the cGAN for mitigating the hard-to-converge challenge in image-to-image (I2I) translation tasks based on classification. The generator side of our proposed model learns to detect the tumor and construct a binary mask, while the discriminator learns to distinguish between ground truth and synthetic masks, driving the generator to produce masks as genuine as possible. The experimental results demonstrate that our BTS-GAN is more efficient and reliable for breast tumor segmentation and outperform other segmentation techniques in terms of the IoU and Dice coefficient on the publicly available RIDER breast cancer MRI dataset. Our proposed model achieved an average IoU and Dice scores of 77% and 85% respectively.