With the development of science and technology, artificial intelligence has been applied to all aspects of our lives.
As a new subversive technology, AI has great accomplishments in mobile phones, face recognition, Go and other fields.
However, did you know that AI is also involved in the medical field?
"AI+ medical care" has always been the focus of scientists' research.
It can not only reduce the medical burden, but also reduce the occurrence of misdiagnosis and missed diagnosis.
So let's take a look at how AI is used in the medical field.
Artificial intelligence and cancer
1 Primary unknown cancer: cancer without primary origin.
CUP (unknown primary cancer), that is, the primary site of tumor origin, cannot be determined.
In the modern precision medical diagnosis and treatment system, there is a class of patients whose tumor tissue has metastasized when they arrive at the hospital, and the existing detection methods can not find the primary focus of their tumor tissue and can not carry out standardized treatment. These patients have short survival time and low survival rate, so we call them "patients with primary unknown cancer".
Using artificial intelligence to find the origin of tumor
On May 5th, 20021,Faisal Mahmood from Harvard Medical School published a research report in the journal Nature, which showed that scientists developed an artificial intelligence (AI) system, which can accurately find the origin of metastatic tumors by using the histological sections obtained routinely, and at the same time, it can also generate a "differential diagnosis" strategy to diagnose patients with unknown primary cancer. [ 1]
TOAD algorithm
Looking for cancer
Artificial intelligence (AI), especially deep learning (DL), can process a large number of high-dimensional data. In the research of Harvard Medical School, AI can use patient histological sections to find the source of metastatic tumors, and at the same time, it can generate differential diagnosis strategies for patients with unknown primary causes.
This algorithm based on deep learning is called TOAD algorithm, which can identify whether the tumor is primary or metastatic at the same time and predict its origin.
The researchers used about 22,000 tumor pathological sections to train the model, and then detected the TOAD algorithm in 6,500 known primary cases, and analyzed more and more complicated metastatic cancer cases, thus establishing the AI model of primary unknown cancer.
For tumors with known primary focus, the model can accurately identify cancer in 83% of the time, and the diagnosis results are included in the first three prediction results in 96% of the time.
Then, the research team tested the AI model in 365,438+07 cancers with unknown primary focus, and found that the coincidence rate between the AI model and pathologists was 63%, and the coincidence rate of the first three diagnoses was 82%.
AI+ medical treatment
There is still a long way to go in the future
Future artificial intelligence medical care
Three development advantages
Cancer diagnosis:
Reduce the "false positive" of diagnosis
In the aspect of diagnosis, the traditional way is to diagnose by computer-aided detection system (CAD), which requires experts to preprocess and screen the data, and manually define the diagnosis rules and related image features. Because of over-reliance on the standard parameters set by experts for diagnosis, it is easy to have false positives.
AI deep learning algorithm can rely on massive data to learn the diagnosis methods of experts, process images autonomously and diagnose diseases.
Automatic extraction of image features;
Find changes that are difficult to observe with the naked eye
Because it is difficult for experts to identify so many quantitative data by watching movies regularly, AI can process large-scale quantitative data and establish correlation at the same time, and each analysis is repeatable.
For example, benign and malignant nodules in the lungs are highly similar, and it is difficult to accurately distinguish them with the naked eye.
AI can automatically extract the characteristics of image biomarkers, detect tiny nodules that are difficult to be found by human eyes, and reduce false positives, so as to distinguish lung nodules and link tumor risk assessment, differential diagnosis, prognosis prediction and treatment effect.
Tumor monitoring and curative effect prediction;
Assist experts to monitor tumors in real time.
The change of tumor volume is an important evaluation index to monitor the therapeutic effect of tumor. When the tumor volume is obviously reduced, it can be recognized by the naked eye, but it is difficult to be recognized by the naked eye if only the tumor texture changes and the uneven changes inside the tumor.
AI can process MRI/CT images at different time points of treatment, learn and extract the corresponding characteristics of tumor texture and heterogeneity changes, accurately identify the areas of tumor changes, and give the thermal map of tumor internal changes, thus assisting clinical experts to judge the treatment effect. [2]
Three difficulties in the future of AI medical care
Data accessibility:
Lack of data support
A reliable AI model needs a lot of high-quality training data to support it, but it is difficult for many hospitals or research institutions to enjoy these data in order to study confidentiality or patient privacy protection. Data "island phenomenon" is a key problem that puzzles the clinical application of artificial intelligence.
Generalization of the model:
Data is not everything.
Generalization refers to the prediction ability of the model to untrained data, that is, the accuracy of the model obtained from hospital A data in hospital B.
The universality of the model is mainly limited by the consistency of the data itself and the subjectivity of data labeling.
Different camera equipment, lighting conditions and individual differences will affect the consistency of image data, and different testing instruments and reagents will also have a great impact on clinical data.
Explanatory result:
Cannot display parameter operation procedure.
Because the internal decision-making process of AI, especially DL, is covered by thousands of training parameters, the weight and characteristics of AI algorithm can not be explained in practice, so it is difficult for clinicians to fully grasp the working process and specific influencing factors of the model. [3]
Of course, the above research results are only the first step of artificial intelligence-assisted prediction of the origin of cancer by using full-slice images. At present, the application of AI in the field of cancer is still in the initial exploration stage.
I believe that there will be more and more AI medical data to establish algorithms in the future. At the same time, with the accumulation of application experience of AI in different diseases, we expect that the diagnostic level of AI will be greatly improved, and it may be expected to standardize the diagnostic process and improve the current cancer diagnosis strategy in the future.
References:
Article/s 41586-021-03512-4
Yuan Zixu, Xu, et al. Application of artificial intelligence in diagnosis and treatment of malignant tumors [J]. China Journal of Experimental Surgery, 2019,36 (2): 203-207.
Zhu W, Xie L, Han J, et al. Application of deep learning in tumor prognosis prediction [J]. Cancer (Basel), 2020, 12(3): 603. DOI: 10.3390/cancer