Genetic algorithm based attention UNet optimization for breast tumor segmentation

Abstract

As one of the main causes of cancer-related mortality among women worldwide, breast cancer requires better diagnostic techniques that can provide non-invasive, fast, and accurate detection. The World Health Organization (WHO) has a dedicated cancer agency called the International Agency for Research on Cancer (IARC), whose mission is to undertake and coordinate research on cancer causes. Mammography is one of many imaging modalities that is frequently used to find abnormalities. Although automated breast mass segmentation in mammography is vital, the uniform sizes and shapes of tumors make it a difficult process. UNet models have shown a significant segmentation in the medical images. In this study, we propose a prominent genetic algorithm (GA) for the generation of UNet models by selecting the optimal parameters. The experiments involved manually generated architectures, basic UNet model and an attention based UNet, AUNet model with different filter sizes. As a result of the manual approach, the AUNet outperformed the base model and hence the AUNet is considered as the base model for the GA process. The experiments show that the models evolved using GA are simple and are of small architecture. The model yielded a better segmentation of the images and outperformed the manually created UNet models, with dice scores and Intersection over Union (IoU) scores of 91.6% and 89.2%, respectively.