AI is expected to make tumor cells nowhere
AI is expected to make tumor cells nowhere December 20, 2018 Source: Chinese Journal of Science If the tumor cells have just been produced, they can be accurately "squeaked out", which will bring about tremendous changes in the diagnosis and treatment of the tumor. To achieve this, the imaging method must have extremely high sensitivity. Recently, the Institute of Automation of the Chinese Academy of Sciences and the Key Laboratory of Molecular Imaging of the Chinese Academy of Sciences have made breakthroughs in the research of new imaging methods based on artificial intelligence (AI) technology. The researchers have three-dimensional positioning accuracy of mouse intracranial glioma. The 100-micron error of the traditional method has been reduced to ten micrometers, providing a new idea for imaging studies of disease animal models and clinical patients. Related research papers have been published in the journal Light. “Image is not obtained out of thin air, but obtained by imaging equipment. Traditional methods often fail to provide the best image quality. Before human image recognition, in the process of converting imaging signals into images, many key information will be lost, artificial intelligence. Technology can break through this bottleneck.†Wang Kun, the first author of the paper and an associate researcher at the Institute of Automation of the Chinese Academy of Sciences, told the Journal of Chinese Academy of Sciences that the original physical signal was transformed into more accurate, higher resolution and more by establishing a new AI model. High-quality images with less artifacts and higher signal-to-noise ratio, whether it is “human brain†or “machine brainâ€, can better identify, recognize and learn. This is the most essential innovation brought by this research. . a very challenging job Early detection of tumors is not easy, especially for some malignant tumors. The incubation period is as long as 20 years. When the body gives out an alarm, it often reaches the middle and late stages. How to achieve accurate detection of early microscopic tumors, timely observation of certain specific proteins, enzymes and even RNAs produced when tumor cells first appeared has been the direction of scientists' exploration and research. "However, in the real physical world, there are not many imaging media that can provide such high sensitivity." Wang Kun said frankly that the best known high-energy gamma rays and non-radiative photons, but based on gamma-detected radioactivity. Radionuclide imaging is costly and difficult to popularize; optical imaging is inexpensive, but most of them are two-dimensional images, lacking three-dimensional information. "We use artificial intelligence to solve the problem that optical imaging is difficult to quantify in three dimensions." Wang Kun said, "that is, you can see whether there is a tumor with high sensitivity, which kind of molecular type of tumor, and you can know with high precision. Where is the tumor, how large is it." The optical imaging mentioned by Wang Kun refers to bio-luminous tomography, which is an important means of biomedical imaging and is widely used in imaging studies of disease animal models. However, since photons have non-homogeneous high-scattering and high-absorption physical properties in living organisms, it is a challenge to reversely reconstruct the position of the light source (ie, the tumor position) in the living body by detecting the illuminating spot of the animal's body surface. Sex work. Luo Jianwen, a researcher in the Department of Biomedical Engineering at Tsinghua University Medical School, told the Journal of Chinese Science. Previously, optical tomographic reconstruction was mostly based on model-based methods, including positive and inverse problems. The solution of positive problems is generally to use the model of radiation transfer equation or diffusion equation to simulate the propagation process of photons in the organization, and then to obtain the system matrix; most of the solutions to the inverse problem use some optimization methods to obtain the specific information of the light source in the body, such as Position, shape, strength, etc. "However, this model-based approach is bound to be affected by the approximation of the model, resulting in reduced accuracy of reconstruction." Luo Jianwen stressed. It is understood that the two errors of the positive problem and the inverse problem solving are superimposed, eventually leading to optical tomography having an error of several hundred micrometers to 1 millimeter for the three-dimensional positioning of the tumor in the animal. Machine learning brings breakthroughs In order to reduce the error, Wang Kun's team proposed a machine learning-based AI reconstruction: completely abandoned the construction of the forward model to describe the propagation of photons in the organism, by constructing a large number of simulation data sets, determining the spot of the animal's surface on the simulation data. The light source in the body, through the data set, trains the computer to intelligently learn the nonlinear relationship between the surface spot and the internal light source, thereby constructing an AI model suitable for bio-luminous tomography, and finally reconstructing the tumor in the living animal tumor-bearing model. Three-dimensional distribution. "This study is the first to apply the multi-layer perceptron method in machine learning to optical tomographic reconstruction, and proposes its own data set construction method, which realizes a cross-model innovation framework directly from data to result, so that the reconstruction positioning error is reduced to One-tenth of the traditional methods, and this also suggests that artificial intelligence methods can be used to solve optical tomographic reconstruction problems," Luo Jianwen commented. However, Wang Kun emphasized that bio-luminous tomography involves the editing and transformation of tumor cells, so it can only be used on animals and not in humans. However, the AI-based optical 3D reconstruction methods developed by them are generalized. Theoretically, it can be used in imaging techniques of other optical molecular images, such as excitation fluorescence imaging, near-infrared imaging, and the like. Therefore, the method itself has a good clinical transformation application ability. Data collection and analysis face challenges The foundation of machine learning is data, and for biomedical imaging, building big data sets is very difficult. “For example, in our study, we constructed nearly 8,000 mouse models of glioma tumors to train our machine learning model. If you really let biologists build a mouse model of in situ gliomas one by one. It takes a long time and puts huge manpower and financial resources into it, which is very unrealistic." Wang Kun said. “The simulation data we built has reached a very high precision and is a good simulation of realistic tumor animals.†Wang Kun said that they used real glioma mice constructed by biologists to verify the artificial intelligence trained. Whether the model is accurate and reliable, the final result shows that the three-dimensional positioning error of the new artificial intelligence method for glioma is less than 80 microns, while the positioning error of the traditional method is more than 350 microns. However, in actual clinical applications, data collection and analysis is not easy. Luo Jianwen said that the most important thing about machine learning, especially deep learning, is data, including the quality and quantity of data. Currently in the field of medical imaging, although it is relatively easy to collect a large amount of data, these data can be used for modeling after being marked, which is greatly affected by individual differences. Because the diagnosis results of different doctors are different, the quality of the data will be affected, and the network trained by it will have problems. At the same time, Luo Jianwen said that in the series of links such as diagnosis, treatment and prognosis, different doctors also have a great degree of freedom for the qualitative description of some diseases. It is difficult to unify the statement; different brands or even the same brand but different models There are also large differences in the images collected by medical devices. These inconsistent data will affect the analysis of deep learning. "An important factor is the causality and interpretability of the model." Luo Jianwen emphasized that "medicine is closely related to human life, so everything must be justified and there must be a causal inference relationship." However, when doing a machine learning model, it is easy to fall into the trap of directly modeling the correlation. The two factors involved in correlation modeling do not necessarily have a direct causal relationship. The derived model, how to interpret the meaning of the result, is A very difficult thing to deal with." Clinical responsibility In Luo Jianwen's view, deep learning is good at dealing with high-dimensional, sparse signals. Image is a representative form of these signals. Therefore, the application of AI in medical image processing is bound to be a hot spot. “The typical problems of medical image processing include image classification, target detection, image segmentation and image retrieval, which can correspond to some pain points in daily clinical applications or waste labor.†Luo Jianwen suggested that imaging physicians should invest in AI. In the learning and application of technology, AI technology may soon help video doctors to complete part of the work, and also has the potential to improve existing work. However, Zhan Songhua, director of the Department of Radiology of Shuguang Hospital affiliated to Shanghai University of Traditional Chinese Medicine, said from the doctor's point of view that AI certainly has a lot to do in detecting lesions, but it is difficult to deal with it instead of doctors. "Discover the characteristics of the lesion, and then distinguish between normal and abnormal, whether it is inflammation or tumor, and ultimately the doctor will make a diagnosis." Zhan Songhua believes that the direction of AI for biomedical imaging is correct, but more research investment is needed now, and doctors and engineers need to be well integrated. AI people need to listen to clinical voices and understand the actual needs of doctors. In addition, AI to solve the false negative rate is the key, to improve the certainty of the AI ​​machine judgment, so as to save time for the doctor. Puyang Linshi Medical Supplies Co., Ltd. , https://www.linshihealths.com