IR-GAN: improved generative adversarial networks for infrared breast image segmentation

ABSTRACT

With medical imaging and artificial intelligence advancements, breast cancer can now be detected at its earliest stages. Infrared imaging is a promising tool in the fight against breast cancer. Segmenting the breast portions from infrared images is essential in computer vision-based diagnosis systems for breast cancer. Generative Adversarial Networks (GANs) have revolutionised the field of image segmentation. By training two neural networks to work together, GANs can produce highly accurate segmentations of medical images, including those related to breast cancer. This paper presents a method for accurately segmenting the breast region from infrared images using improved GANs. The use of reconstruction loss to generate accurate segmentation masks using a U-net-based generator is proposed in this study. The PathchGAN discriminator was trained to distinguish between the generated and ground truth images. The generator was updated through feature reconstruction loss and adversarial loss using the knowledge acquired by the discriminator. Additionally, a novel activation function called ‘AR-erf’ was introduced in this paper for generating accurate segmentation results. The proposed model was evaluated on three different data sets, yielding a dice score of 0.94 and a mean intersection of union of 0.932. This can be compared to other reported techniques in the literature..

KEYWORDS:

• Generative adversarial networks
• infrared image processing
• medical image segmentation

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

1. The data that support the findings of this study are openly available in the‘Database for Mastology Research’ at https://visual.ic.uff.br/dmi/.

2. The data that support the findings of this study are openly available in the ‘Thermographic mammary images database for breast cancer research’ at https://github.com/Biomedical-Computing-UFPE/MammoTherm.

Notes

1. *Appendix A and B shows a larger set of Segmented breast images. Appendix C shows how the images are inaccurately segmented when using the regular Re-LU activation function.