Deep network lung texture recognition method combined with multi-scale attention

Abstract

The invention discloses a deep network lung texture recognition method combining multi-scale attention, and belongs to the field of image processing and computer vision. In order to accurately identify typical textures of diffuse lung diseases in a computed tomography (CT) image of a lung, the invention provides a lung diffuse lung disease identification method. By designing a unique attention mechanism module and a multi-scale feature fusion module, a multi-scale and attention combined deep convolutional neural network is constructed, and high-precision automatic identification of typical textures of diffuse lung diseases is realized. In addition, the proposed network structure is clear, easy to construct and easy to implement.

Description

technical field [0001] The invention belongs to the fields of medical image processing and computer vision, and in particular relates to a deep network lung texture recognition method combined with multi-scale attention. Background technique [0002] Diffuse lung disease refers to the general term for abnormal texture of the lung interstitium in the lung area caused by factors such as inflammation or injury. CT images are often used for the detection of such diseases because they can clearly present the state of lung tissue. However, due to factors such as the large number of collected images and the complexity of lung textures, it is difficult even for experienced radiologists to accurately identify different types of lung textures, resulting in missed diagnosis and misdiagnosis. Therefore, it is necessary to establish a computer-aided diagnosis (ComputerAided Diagnosis, CAD) system to assist radiologists in performing accurate and efficient automatic diagnosis of lung tex...

AI Technical Summary

Problems solved by technology

First, lung textures in CT images present two radiological features, grayscale and scale information. Currently, most CNNs used for lung texture recognition only use grayscale feature information, and have not yet focused on learning the scale features contained in lung textures. information, so it is necessary to design and use a mechanism to enable CNN to learn multi-scale feature information of lung texture

Second, the CNN parameters currently used for lung texture recognition are generally large in size. The feature maps learned by the convolutional layer in CNN have redundant information, which affects the final recognition accuracy. It is necessary to design and use a mechanism to automatically screen for the recognition task. Feature maps, while automatically suppressing feature maps that are weakly related to the recognition task to reduce the impact of redundant information in the feature map and improve the final recognition accuracy

Images