What do we have to know about PET/CT?
PET/CT is a nuclear medicine device that perfectly integrates PET and CT systems. PET provides detailed molecular information on the function and metabolism of the lesion, while CT provides precise anatomical positioning of the lesion, and one-time imaging can obtain tomographic images of the whole body. Compared with independent PET and CT, PET/CT can significantly improve the sensitivity, accuracy, specificity and positioning accuracy of diagnosis. It can understand the overall condition of the whole body at a glance, and achieve the purpose of early detection of lesions and diagnosis of diseases. It is mainly used for early detection and diagnosis of major diseases in tumor, brain and heart fields.
According to the big data of the Global Cancer Annual Report released in 2018, there are an estimated 18.1 million new cancer cases and 9.6 million cancer deaths worldwide. There is 1 cancer patient in every 65 people in our country, which is the leading cause of death. According to statistics from the World Health Organization, the current cure rates and survival rates of various treatment methods are not satisfactory to people. The main reasons are that the diagnosis is too late, the staging is inaccurate, and the treatment is incomplete. Because PET/CT can observe changes in cell metabolism in the body, it is possible to clarify the nature of the tumor's primary tumor (differentiation of benign and malignant tumors, tumor staging and grading) before structural and morphological changes, and whether there are systemic metastatic lesions (systemic conditions) Evaluation), how the effect is (a few days or even hours after radiotherapy can observe the therapeutic effect of the tumor, adjust the treatment plan in time, and radically cure the incompletely treated lesion), etc. In addition, PET/CT has unique advantages in the localization of brain epilepsy lesions before surgery, the identification of radiation necrosis and recurrence after tumor treatment, the classification of brain tumor malignancy, and neurological diseases.
However, the radiation brought by PET/CT scans often makes people "distracted". The amount of radiation taken in a PET/CT whole body examination is about 7.5mSv. what is this concept? In nature, humans receive approximately 2.4mSv of natural radiation each year, so the dose of PET/CT examination cannot be ignored. In response to the radiation dose problems of radiation and radiation caused by CT and injected PET radiopharmaceuticals in PET/CT scans, the World Health Organization, the International Radiological Commission and the International Medical Physics Organization have formulated medical exposure quality assurance and dose control standards, and strongly advocated Radiation exposure should follow the ALARA (As Low As Reasonably Achievable) principle of practical legitimacy and optimal protection. It is expected that the best diagnostic images can be obtained with the smallest radiation and radiation dose, while further reducing the cost of PET/CT inspections and reducing scanning time.
However, reducing the injected radiotracer will amplify Poisson noise, which will affect the image quality, lesion detection and quantitative accuracy of PET. In low-dose imaging, many key information will be submerged under the increased noise level. By redesigning/optimizing the reconstruction algorithm of low-dose scanning, the best trade-off between noise level and signal convergence can be achieved. In order to solve the above-mentioned challenges, many algorithms and technologies have been proposed, which can be mainly divided into traditional algorithms and deep learning algorithms. Among them, the traditional algorithms mainly include post-reconstruction processing/filtering algorithms, anatomical guidance algorithms, statistical modeling in the iterative reconstruction process, and noise removal and partial volume effect correction under the guidance of MRI. Although these methods try to minimize noise and quantitative errors, there are still problems with spatial resolution loss and excessive smoothing.
Deep learning algorithms have recognized capabilities in solving complex inverse problems, such as image reconstruction from projections. The image reconstruction process of CT, PET and SPECT using deep learning technology has roughly the same methods. There are currently four main strategies: The first method is the image-to-image learning process, that is, the image-to-image learning process is performed in the image space. Image conversion, training a network model to improve the image quality of the reconstructed image through denoising and super-resolution modeling. The second method is the sinogram-to-sinogram learning process, that is, training a deep learning model in the projection domain to improve the image quality of the sinogram to avoid sensitivity and dependence on the reconstruction algorithm. The third method is the sinogram-to-image learning process, that is, learning the nonlinear mapping relationship between the projection domain and the image domain through the network model, completely removing the traditional reconstruction algorithm, and generating the image in one step. The fourth method can be called hybrid domain learning. By fusing the reconstruction algorithm and deep learning, the network model is trained in the projection domain and the image domain at the same time to realize the optimal solution of the image reconstruction problem.
The current industry generally uses low-dose PET imaging algorithms in the image domain, that is, after the PET/CT equipment outputs the image, the image quality is improved through image post-processing. Due to the large noise in low-dose PET images, these noises conceal many fine structures in PET images. This technical route usually leads to image artifacts, quantitative errors, and loss of fine structures. Traditional PET imaging has lost a lot of information in the reconstruction process. It is extremely difficult to recover the lost information only by processing the image in the later stage, and it is difficult to improve the final image quality. In order to solve the problem of image quality from the source, some medical and laboratories have innovatively developed deep learning algorithms based on PET raw data and reconstructed images (the fourth type of hybrid domain learning). The algorithm deeply embeds AI into the PET image reconstruction process, and uses deep learning to mine the information in the original data. By combining the physical model of PET reconstruction, the processing object is directly advanced to the original data inside the imaging device, and the reconstruction algorithm is assisted to improve the quality of the reconstructed image, which greatly reduces the loss of effective information, so as to obtain clearer PET images and stronger The ability to detect small lesions.
