Semantic segmentation of breast lesion using deep learning

Abstract

Objectives Although breast ultrasound imaging is powerful and effective tool to detect breast lesions and have been widely performed worldwide, it is an operator-dependent test, hence the accuracy for detection and diagnosis of breast lesions depend on the operator. We develop a computer-aided segmentation system for breast lesions in ultrasound images using deep learning. Methods A data set containing 215 ultrasound images were collected from our institute. Ultrasound images were cropped at 512 × 512 pixel images automatically. The breast tissue area within each image was extracted and cropped to make a masking image. The U-Net algorithm was programmed to train the images with deep learning. The original ultrasound images were input as the input images, and the corresponding masking images were input as the output images to extract breast tissue area. DICE coefficient was used to evaluate its accuracy of extraction. Also, the breast lesion was cropped to make masking image and these images were also trained by U-Net to extract the area of the lesion. Results DICE coefficient achieved over 90% of accuracy to extract the area of breast tissue. The breast lesion was also successfully extracted from the images. Conclusions We achieved accurate segmentation of the area of breast tissue and breast lesions. Deep learning could potentially help detecting and diagnosing the breast cancer, improving accuracy and productivity of detecting and diagnosing breast cancer.