U-Net and SegNet performances on lesion segmentation of breast ultrasonography images

Abstract

Purpose: Screening is the predominant strategy for the early detection of breast cancer. However, image analysis depends on the experience of the radiologist, inserting subjective factors in the evaluation of findings. Most biopsies performed on women with suspicious lesions on breast imaging tests show benignity. This medical technique is invasive and consumes time and resources; therefore, there is a demand to reduce these unnecessary procedures. Computer-aided diagnostics has been increasingly used as a second-reader tool lately, decreasing the diagnostic uncertainty of specialists. Based on this fact, the development of such systems is crucial and can be accomplished by the refinement of its steps, i.e., more accurate segmentation and classification. Considering this scenario, the present work describes the implementation of two well-known convolutional neural networks, U-Net and SegNet, for the segmentation of lesions found in breast ultrasonography. Defining which architecture is most appropriate for this task can ultimately help in reducing the number of biopsies performed. Methods: Convolutional neural networks are used in segmentation by classifying each pixel in an image, based on self-trained weights. Using a dataset of 2054 images, obtained in partnership with the National Cancer Institute, we compared the automatic segmentation performed by the networks with a manual segmentation made by a specialist. Results: Among the two proposed architectures, U-Net obtained better results in this task, obtaining a dice coefficient of 86.3%, and took 68.3% less training time than the SegNet, which achieved a dice score of 81.1%. Conclusion: The two networks can segment ultrasound images with useful accuracy depending on their configuration, with U-Net being faster to train and more accurate.