Disease diagnosis system and method using neural network and non-local blocks for segmentation

Abstract

Disclosed are a disease diagnosis system and a method thereof that can segment a region in which a disease is present in an image of a biological tissue by learning through a neural network and using the learned neural network and a non-local block. According to one aspect of the present invention, there is provided a disease diagnosis system implemented in a system comprising a processor and a storage device for storing a neural network, using a slide as a biological image and the neural network, the disease diagnosis system comprising a patch level segmentation neural network, the patch level division neural network receives the patch as an input layer and specifies a region in which a disease is present in the patch for each predetermined patch that divides the slide into a predetermined size, and includes: a patch level classification neural network that receives the patch as an input layer and classifies a patch level of the patch as an input layer; outputting a patch level classification result related to whether the disease exists in the patches or not; and a patch level segmentation architecture for receiving feature maps respectively generated in two or more feature map extraction layers in a hidden layer included in the patch level classification neural network, specifying an area in which a disease exists in the patch, the patch level segmentation architecture comprising: a non-local correlation calculation sub-architecture, comprising non-local correlation calculation nodes respectively corresponding to the two or more feature extraction layers, and the non-local correlation calculation nodes respectively carry out a non-local correlation calculation process on the feature maps input from the feature extraction layers corresponding to the non-local correlation calculation nodes. The non-local correlation calculation process is a convolution process, a non-local block process or a parallel convolution and non-local block process; and segmenting the sub-architecture, and specifying a disease region in the patch on the basis of the result generated in the non-local correlation calculation sub-architecture.

Description

technical field [0001] The invention relates to a disease diagnosis system and method using neural network. In more detail, it relates to a disease diagnosis system and method thereof capable of segmenting a region where a disease exists in an image of a biological tissue by performing learning through a neural network and utilizing the learned neural network and a non-local block. Background technique [0002] One of the major businesses performed in pathology or the department of pathology is diagnosis by reading biological images of patients to judge the status or symptoms of a particular disease. This diagnosis is a method that has long relied on the experience and knowledge of skilled medical personnel. [0003] In recent years, thanks to the development of machine learning, businesses such as image recognition or classification are being actively attempted to be automated by computer systems. In particular, attempts are being made to automate diagnosis by skilled med...

AI Technical Summary

Problems solved by technology

That is, pathologists usually look around and make a diagnosis, but most existing deep learning networks cannot reflect these

Images