Research On Pulmonary Medical Image-to-Image Translation And Its Aided Diagnosis Techniques Based On Generative Adversarial Networks

According to the authoritative d release of IDC,an international data company,the total value of the global artificial intelligence application market will reach 127 billion dollars by 2025.The medical industry will account for 1/5 of the total market.In 2017,the State Council officially issued the "New Generation Artificial Intelligence Development Plan" and pointed out that in the medical field related to the needs of people’s livelihood,a fast and accurate intelligent medical system should be established.In the medical field,artificial intelligence-assisted diagnosis is mainly based on medical images.Manual analysis of medical images has a high misdiagnosis rate and doctors are prone to fatigue.Artificial intelligence based on deep learning can process medical images in large quantities,reduce the time for doctors to read images,and work without fatigue and without interruption.It has been proven to greatly improve the speed of medical diagnosis and the accuracy of disease identification.Deep learning models generally rely on a large amount of training data.However,the available,high-quality and well-labeled medical imaging datasets are insufficient.The medical images are uneven and the medical images with diseases are even scarce.It cannot guarantee the generalization performance of the auxiliary medical diagnosis model.The main reasons for the lack of available medical images is the need to protect patient privacy.Professional doctors have heavy clinical and scientific tasks,and have no time to complete the labeling work.The medical image-to-image translation based on Generative Adversarial Network(GAN)is an important method for artificially synthesizing the required medical images.It can effectively provide synthetic images.It is of great significance to alleviate the scarcity of medical image data and improve the speed and accuracy of medical assisted diagnosis.It has become a research hotspot in the current medical synthetic image field.Relevant literatures show that the image-to-image translation method based on generative adversarial networks still has some problems,such as insufficient utilization of existing data,poor attention and insufficient ability of desired translation,and poor quality of generated images,which limit the use of synthetic images in the medical aided diagnosis network based on deep learning.In view of the above problems,this paper conducts research on the medical image-to-image translation based on GAN,and further verifies the effectiveness of the expansion based on synthetic image data for medical assisted diagnosis.The main innovations are as follows,1)Aiming at the problems that the medical image-to-image translation model does not make full use of the existing data and pays little attention to desired transfer branches,a multi-domain medical image-to-image translation algorithm with desired transfer branches based on GAN is proposed.After analyzing the imbalance of medical images,this algorithm selects key target image domains and establishes desired transfer branches.It completes the conversion between multi-domain medical images using a single generator.Also,it makes full use of existing images and labels as monitoring information to ensure the attention ability of the multi-domain medical image-to-image translation model and quality of the synthetic images.Finally,the conversion of normal images without disease,viral pneumonia and bacterial pneumonia images was completed in the chest X-ray image dataset.The score of GAN-test and GAN-train of the synthetic images reaches 92.188%and 85.069%.This model has a notable improvement over other generative models in terms of authenticity and diversity.2)For the problem of poor quality of synthetic images,a multi-domain medical image-to-image translation model with dual feature constraints based on GAN is proposed.Under the premise of ensuring desired transfer branches and attention performance of the model,this algorithm extracts the middle layer features of the input image by discriminator.The feature consistency between the reconstructed image and source domain image and the feature consistency between the generated image and target domain image constraints the image translation process to ensure the authenticity and diversity of the synthetic images.It was verified on the chest X-ray image dataset and achieved good results.The normal disease-free image and the two types of pneumonia images were converted to each other.GAN-test score increases to 94.792%,and the GAN-train score is 85.069%.The background smoothness of synthetic image is reduced obviously and the synthetic image details are rich.3)A pneumonia-assisted diagnostic model based on synthetic images data argumentation is proposed.Synthetic chest X-ray images and original images are merged as training datasets to study the performance of the aided diagnostic model in the classification of bacterial pneumonia and viral pneumonia.The pneumonia diagnostic accuracy of model reaches 93.8%,which is 2.1%higher than the model trained using only original images.The sensitivity of the model is 96.69%,increasing about 7.1%.And the specificity is 89.20%.The area under the ROC curve increases from 94.00%to 96.25%.The results verify the effectiveness of the synthetic images data argumentation generated by the medical image-to-image translation model for artificial intelligence-assisted diagnosis,and help doctors to provide targeted diagnosis measures for pneumonia in the next step.