Comparison evaluation of unet-based models with noise augmentation for breast cancer segmentation on ultrasound image

Abstract

This paper investigates the application of convolutional neural networks (CNNs), particularly the UNet model architecture, to improve the accuracy of breast cancer tumor segmentation in ultrasound images. Accurate identification of breast cancer is essential for effective patient treatment. However, ultrasound images, often contain noise and artifacts, which can complicate the task of tumor segmentation. Therefore, to highlight the most robust architecture, modifications were made to the original set, including the addition of noise and fuzziness. In this study, a comparative study of five different variants of UNet models (UNet, Attention UNet, UNet++, Dense Inception UNet and Residual UNet) was conducted on a diverse set of ultrasound images with different breast tumors. Using consistent training methods and techniques of augmentation and adding noise to the data, an improvement in segmentation accuracy was highlighted when using the Dense Inception UNet architecture. The results have potential practical applications in the field of medical diagnosis and can assist medical professionals in automatic tumor segmentation in breast cancer ultrasound images. This study highlights the improvement of segmentation accuracy by introducing dense induction into the UNet architecture. Importantly, the Dice coefficient, a key segmentation metric, improved markedly, increasing from 0.973 to 0.976 after data augmentation. The results of the study offer promise to the medical community by offering a more accurate and reliable approach to segmenting breast cancer lesions on ultrasound images. The findings can be implemented in clinical practice to assist radiologists in early cancer diagnosis.