A Deep Learning Neural Network Model System for Multimodal Image Synthesis

Abstract

The invention relates to a deep learning neural network model system for multimodal image synthesis, including a multi-resolution residual deep neural network formed by combining a residual deep neural network (RDNN) with a multi-resolution optimization strategy; RDNN includes a A convolution layer, B drop-off layer, C batch normalization (Batch Normalization) layer and D long-term residual connections; among them, the convolution layer is used to extract image features; the drop-off layer is used to avoid network overfitting ; The batch normalization layer is used to normalize the input of the corresponding convolution kernel; the long-term residual connection is used to preserve the structural information in the input image; both sides of each shedding layer are provided with convolutional layers, and each shedding layer The convolution layers adjacent to both sides are connected; a convolution layer is set between each drop-off layer and the batch normalization layer; the drop-off layer, the convolution layer and the batch normalization layer are connected in sequence; each One end of a long-term residual connection is connected between the convolutional layer and the batch normalization layer; the other end is connected between another set of convolutional layers and the batch normalization layer.

Description

technical field [0001] The invention relates to the technical field of medical image processing and guided treatment, in particular to a deep learning neural network model system for multimodal image synthesis. Background technique [0002] Adaptive Radio Therapy (ART) technology based on CBCT images can increase the irradiation of radiation to the tumor, and at the same time, protect the organs at risk near the tumor. However, the HU accuracy of CBCT images in the prior art is low, the resolution of soft tissues is low, and there are serious artifacts. Therefore, in order to realize adaptive radiotherapy (ART), synthetic CT (sCT) images with high HU accuracy and structural fidelity based on CBCT images are firstly required. [0003] U-Net and some other deep learning networks have been widely used in the task of sCT image generation. The study showed that the HU accuracy of the generated sCT images was significantly improved, and dose calculations could be performed based...

AI Technical Summary

Problems solved by technology

Second, the above network lacks constraints to improve the structural fidelity of sCT images

Images