Internal diseases can be decomposed into a series of abnormal gene expression patterns, which can reflect different properties at different levels, such as tissue changes at the cellular level (hyperplasia, atrophy, and large chemical fertilizers), from the bottom to the top, and various pathological manifestations at the body level. We can understand it in terms of macroscopic mechanisms at different levels, such as abnormalities in the immune system. If we want to treat, we don't just want to target objects that express specificity such as enzymes, hormones, etc., and maybe we need to understand the disease in a different way of thinking, such as the idea of the Fourier transform (a combination of continuous signal changes in the time domain into the frequency domain). The more metaphorical view is that it is broken down into a system of meridian acupuncture points (tender points and so on may be immovable points), which is actually a kind of essentialism (the so-called essence is to point directly to the cause), and there is a kind of high-dimensional objects that can accept complex changes at the lower dimensional level (similar to calculus), and we can play a therapeutic role by processing these high-dimensional objects. For example, in traditional Chinese medicine, the concept of sick qi, evil qi and so on, if the meridians are smooth, then the disease can be eliminated. This view of meridian obstruction is to extract the basic objects in the frequency domain, and we are able to determine the treatment plan through a selective combination of these linear independent substrates. It is possible that this explanation is the mathematical basis of the theory of TCM.

In fact, the prescription is a corresponding solution, which can meet the complex conditions of the human body to construct a differential equation model. In fact, we regard specific Chinese medicine as a substrate, and the specific dose is the weighting of these substrates, and their combination is a specific vector of high-dimensional space. And we have so many traditional Chinese medicines as a base, it is impossible to search for possible combinations, we need a variety of a priori IQ to search the scope of our search, such as we divide the medicinal properties (similar to dividing yin and yang), and then further classify on this basis, and then combine them (medicinal properties to neutralize each other, etc., monarchs and ministers, etc.), and store meaningful combinations (various mature prescriptions, using the idea of greedy algorithms). The specific ratio can be understood as the determination of parameters, and we can use gradient descent method and so on to determine. We can express both the specific basic and the treatment as a certain function, f(n)=∑anx^n, and we can approximate the possible actual situation with arbitrary precision.

Of course, theoretically we need to do a detailed decomposition to get more accurate results, but it is also very important for us to operate at a sufficiently high-dimensional level, and in fact our doctors are functionally more deterministic at these macro levels, just like the mapping construction of the hidden layer of machine learning/deep learning. For example, various symptoms are common in those diseases, which diseases are typical and atypical symptoms, which are high-dimensional level data sets, and our medical learning is through the training of these training sets to build certain functions in the brain, which can achieve various functions such as diagnosis and so on. After all, if you think about doctors with object-oriented thinking, you can input data such as symptoms, consultations, physical examinations, laboratory tests, and imaging tests to obtain a diagnosis of a specific disease as an output function.

At this high-dimensional level, we can comprehensively consider the synthesis of various data, continuously improve the possible probability of different diseases in the way of Bayesian operation, and then check according to the order of the probability. In specific clinical practice, doctors use a variety of information with high evidence strength to guide the diagnosis (feature finding and linear combination), such as various characteristic lesions can quickly arrive at the diagnosis of classical diseases. Moreover, the occurrence and development of the disease is a dynamic process, which can decompose the specific symptoms into a lower level of pathological pathogenesis, and even to the anatomical, histological and molecular level mechanisms, we need to consider that the disease may be the result of the comprehensive action of various underlying factors (understood as a certain vector of linear combination, different vectors corresponding to different combinations can correspond to different diseases), and by considering multiple details, we can finally converge to the fixed point, that is, we think the etiology, so as to be able to prescribe the right medicine in a targeted manner, play the role of four or two thousand pounds, so that others can restore health. This idea of the etiology of the immobility point is a classic application of essentialism, in which we believe that it is possible to achieve this diagnosis by means of a computer that converges to a specific value. This reduces the need for doctors to use different drugs and perform different tests to narrow down the range of possible diseases, i.e., to reduce the cost of searching for convergent results. And this exclusion is the general differential diagnosis idea.

Laboratory tests can provide a lot of data that can be quantified, and theoretically, as long as there are enough indicators, we can classify all patients to a certain extent, and then implement individualized medicine according to the methods (evidence-based medicine) that a specific classification of people can benefit from at the group level. Of course, there is a risk that this will be too detailed, so that you can't see the forest for the trees. How can we grasp the more important essence, we hope to be able to approximate the higher dimensional level through the infinite accumulation of the lower dimensional levels, as the fundamental theorem of calculus does. Predictably, this is also quite computationally staggering. Through the measurement of indicators and the meaning of the learning system based on experience, we can finally integrate and use certain algorithms such as classification algorithms to simulate the diagnosis of doctors.

Medical imaging is a comprehensive diagnosis and treatment discipline based on medical imaging, integrating X-ray, computed tomography (CT), magnetic resonance imaging (MRI), digital subtraction (DSA), positron tomography (PET), nuclear medicine, ultrasound (US), radiotherapy and interventional therapy. Multi-dimensional data can help us have a more accurate diagnosis of diseases. The focus is on finding specific changes, i.e., lesions corresponding to various diseases, and we can obtain meaningful information by analyzing these differences. Combined with the above existing diagnostic thinking, we can greatly save the scale of the calculation, and we can constantly classify (of course, it is also possible to be led astray by the wrong experience of doctors, which is a trade-off, but learning experience is always a good entry point). We believe that the power of data analysis lies in the ability to analyze multiple indicators at the same time (Bayesian inference), and that human attention resources are limited and can only perform calculations on a limited number of key indicators, so we pay great attention to characteristic changes (such as morphological changes), after all, we can make meaningful inferences with greater accuracy. By identifying the basic lesions of the system, we can further construct them to obtain complex disease diagnoses (our ideal is that in the future, we will not need a final disease diagnosis, but will obtain an individualized treatment plan through the input of various indicators. Classification conforms to human thinking habits, but computers are able to process a broader range of metrics and then abstract out finite classifications, which is a learning process (pruning strategies for searching for space). We need to not only identify the lesions, but also understand the mechanisms of pathophysiological alterations).

In the process of ultrasound imaging, we need to consider not only the sensitive scattering of different tissues to ultrasound, but also the attenuation during the ultrasound transmission process. We provide feedback on the impact of these external stimuli on the internal tissues to imaging, which is essentially a construction, data analysis, and the confidence that the results can correspond to the real situation one by one, of course, this is the ideal situation, and many times the imaging has artifacts and so on, but this is the premise of our continuous advancement of technology.

Imaging is understood as a signal processing process, and through the process of sampling, detection, analysis, and image formation, we can form macroscopic information that we can understand. From one-dimensional A-type linearity, to two-dimensional flat images, to three-dimensional morphology, and four-dimensional time variation.

Pathology can become the gold standard, on the one hand, because morphological changes are relatively high-dimensional changes, which are terminal states caused by a series of etiologies, and finding their characteristic changes can describe the relationship between many objects at the same time. At the same time, it is also because of its large number of materials, there are always specific sections that can have the highest level of correlation with specific diseases, that is, one-to-one correspondence, so that changes at the organism level can be inferred from changes at the cellular level. Of course, tissue from the vicinity of the lesion needs to be taken, as it has a greater relevance.

Our question is, how can machine learning algorithms be combined to train human image-based disease diagnosis? We need to have large-scale labeled data, which can be used in cooperation with hospitals to train algorithms using images diagnosed by experts in databases; At the same time, we need to carry out accurate lesion identification, that is, image segmentation, to find abnormal changes, so as to analyze on this basis; It is foreseeable that the complexity of the calculation is huge, and we need to use a series of optimization strategies and approximation algorithms to arrive at a preliminary prototype, and then continuously adjust and continuously improve various quantitative indicators such as precision, recall, etc., until it exceeds the diagnostic level of human experts. For example, Dermatologist-level classification of skin cancer with deep neural networks uses deep learning algorithms to identify skin cancer. It can be accurately classified in a subdivided field. Finally, on the basis of the accurate classification of this diagnosis, we can further explore various individualized treatment methods, so as to achieve our goal of precision medicine.