Nuclei Segmentation utilizing MSCNN and Guide image

Abstract

Recently, image processing technology has been applied to various fields and to be beneficial for human life. For example, image processing technology plays an auxiliary role in medical examinations and aids accurate diagnosis. A representative example is the detection of a nucleus in a medical image. The reason for extracting nuclei from medical images is to identify the presence or absence of disease and to obtain clinical information by identifying shapes and numbers of nuclei. Numerous studies have been conducted recently to improve image processing performance through the application of deep neural networks. In this thesis, a robust method is proposed for detecting nucleus regions in medical images (captured by various imaging equipments) using deep neural networks. The network architecture proposed consists of a parallel network that extracts features of different sizes in an image and a classification network that classifies the nucleus of each pixel in units of different size. Particularly, parallel networks use atrous convolution instead of pooling (widely used for feature extraction and reception area amplification) to adjust the size of receptive field and to remove the resolution restoration process. In addition, a residual network is configured to enable more accurate partitioning. The input of the network consists of preprocessed medical images and guide images containing approximate information of the nucleus region. The proposed method has 6.42% higher mIOU and 1.73 times faster than Deeplab-v1 with pooling removed.