Breast Lesion Segmentation in Ultrasound Images Using Deep Convolutional Neural Networks

Abstract

On account of the painful impact of mammography, ultrasound imaging has gained wide acceptance in screening and diagnosis of breast lesions. But the existence of speckle noise and low resolution of ultrasound images limits the accuracy in some of the significant tasks like automated localization and segmentation of any diseased prone region. Therefore, to aid in computer-aided diagnoses, and to address the complex structure of breast lesions in ultrasound images, an efficient segmentation method can clinically assist in the treatment of patients by locating potential areas of interest and quantitative measurements of clinical information. Till date, most of the current methods for breast lesion segmentation from ultrasound imaging depends on manual delineation methods that are acutely tedious and impressionistic and may lead to human errors. As a consequence, an automated method for breast lesion segmentation from ultrasound images is of high demand from the radiologists now a day. The overarching objective of the work is to establish a light weighted convolutional neural network architecture encuraged by two popular networks (U-Net and ResNet), on 163 ultrasound images with size 760 x 570 pixels from patient with different types of diseases like benign cyst, malignant invasive ducal carcinoma (IDC), to automatically segment the breast lesion. The proposed architecture used a feature extraction unit at the beginning from a stack of convolutional layers to extract features from the input dataset. The final prediction of the architecture has been made after feeding the features from the U-Net and modified ResNet. This approach results in significant improvement in terms of accuracy, dice coefficient, mean-IoU, recall and precision with respect to some of the state-of-the-art architectures. Our comprehensive evaluation, model analysis, and antique insights can be extended to the development of the Computer-Aided Diagnostic (CAD) tool to assist the Ultrasound image reading process for the detection of breast lesion in the near future.