Research On PET Image Quality Improvement And Model Compression Methods Based On Deep Learning

With the rapid development of medical imaging technology,medical images can intuitively provide information such as lesion location,structure and function,which is an important means and diagnostic basis for the diagnosis of various diseases.And the continuous development of deep learning technology and the significant enhancement of computational resources have made computer technology-based computer-aided diagnosis(CAD)a research hotspot in the fields of medicine and computer,and widely used in the analysis and processing of medical images.Positron emission tomography(PET),an advanced imaging technique in nuclear medicine that reflects the metabolic distribution of fluorodeoxyglucose in humans,is currently an important means of early tumor screening and diagnosis.Limited by the principle of PET imaging,the sensitivity of conventional short-axis PET scanner is low,resulting in low signal-to-noise ratio of image.The uEXPLORER is an advanced total-body PET imaging scanner with 1940 mm field of view(FOV).It has ultra-high sensitivity and can obtain image quality as the gold standard.In order to improve the image quality of short-axis PET scanner,based on the high-quality and short-axis data collected by uEXPLORER,combined with the field knowledge of PET imaging principle,computer vision and deep learning,this paper constructs a method to improve the short-axis PET image quality based on deep learning,and studies an efficient lightweight compression model.The main work of this paper is as follows:1)Aiming at the problems of low imaging quality of conventional short-axis PET scanner in a short time,a PET image quality improvement model based on cycleconsistent adversarial networks(CycleGAN)was proposed.The introduces attention mechanism is introduced into the generator to realize the attention to the channel and spatial representative features.Combined with the supervised paired data learning scheme,the spatial consistency between the generated and the real image is realized.The method proposed in this paper is compared with the traditional gray-scale based and learning based methods.The experimental results show that the images generated by the proposed method have the best SSIM and NRMSE indexes performance in the three beds,which are(0.219±0.09,0.950±0.03),(0.222±0.06,0.927±0.02)and(0.257±0.07,0.958±0.02),respectively,and have been recognized by imaging doctors.This method can obtain good image quality and good adaptability in PET scanners with 320 mm and 250 mm FOV.The image quality improvement model can effectively improve the imaging quality of conventional short-axis PET scanner and give full play to its clinical value.2)Aiming at the problems of large scale,difficult deployment and efficient operation of deep learning image quality improvement model,a model compression method based on knowledge distillation is proposed.Using knowledge distillation technology to extract the intermediate features of the original teacher model,transfer the knowledge to the corresponding compressed student model,and use neural architecture search technology to find an efficient structural design requiring lower calculation cost and less parameters in the student model,so as to realize the model lightweight and effectively improve the image quality of short-axis PET scanner.This study compares the performance,parameters and amount of calculation of the model before and after compression.The experimental results show that,while maintaining the performance of the model,the model parameters and computation are 4.5%and 6.8%of the original respectively.Efficient lightweight compression model can satisfy the high real-time and concurrent use of CAD system in clinical,making it play a greater clinical role.