AI pathological diagnosis interpretation scheme or solution to the key difficulties of AI CFDA application for approval

Deep convolution neural network (CNNs) has been proved to be a kind of technology that can assist biomedical image diagnosis in practice, and has been widely used in lung nodule, fundus and other radiation image recognitionRecently, AI research in the field of pathology has also made new progress< br / > in May 2019, the paper "pathogist level interpretable whole slide cancer diagnosis with deep learning" by Yang Lin team in China was included in "nature machine intelligence", which proposed a scheme for AI pathological diagnosis and interpretation< br / > in the experiment described in this paper, the researchers used AI technology to analyze and process pathological sections, and provided the basis for AI analysisThis is the world's first monograph on the interpretability of artificial intelligence in pathological image analysis published in nature< br / > through the method designed by the experiment, AI began to "understand" the doctor's logic, and tried to imitate the human doctor to give the diagnosis basisIn this regard, the arterial network interviewed Professor Yang Lin, the corresponding author of the paper, and combined with the content of the paper, tried to sort out the logic of the paper and the profound value behind it< br / > pathology department promotes the development of scientific research < br / > pathology department is called "the foundation of medicine" by William Osler, the "father of modern medicine", and pathologist is considered as doctor's doctorThe gold content of pathology department is self-evidentThe accuracy of its diagnosis directly affects the health and fate of patients< br / > however, according to 2015 data from the national health and Family Planning Commission, there are only 9841 qualified pathologists in ChinaThe ratio of this figure to the total population of our country is about 1:140000, and the ratio to registered doctors is about 1:250Simply put: every pathologist has undertaken 5-10 times of the routine workloadMany pathologists are overloaded with increasingly complex and high-intensity work, and misdiagnosis and missed diagnosis occur frequently< br / > the factors that restrict the development of pathologists' resources are not only the huge workload, poor working environment, low income and treatment, long training cycle and other factors that seriously affect the pathologists' teaching staffThe new power of pathologists is in short supplyThe emergence of AI technology may solve this problemAI supported by deep learning can process medical images in a rapid and standardized way, sketch and render suspicious images, and make suggestions in structured language< br / > these tasks consume a lot of energy and are highly repetitive, while AI is not subject to the nature of workPractice has proved that with the help of AI, pathologists can not only improve the efficiency of diagnosis and reduce the workload, but also improve the work intensity, improve the working environment of pathologists, and ultimately reduce the rate of misdiagnosis and missed diagnosis< br / > pain points really promote the development of scientific research, but when AI assisted diagnosis is really applied, various problems follow< br / > the most clear and difficult questions to answer are the following two questions: how does AI complete the interpretation? Does it have a basis for slice analysis? In fact, if this problem is not solved, it is difficult for pathologists and CFDA regulators to recognize the interpretation results of AI - probability cloud is not a reasonable basisIn view of this, Yang Lin team started this study to solve the feasibility and interpretability of AI pathological diagnosis< br / > under the experimental conditions, AI can greatly improve the accuracy of CAD < br / > in order to explore the interpretable problems in the process of AI assisted diagnosis, the research team took the pathological section of bladder cancer patients as the research object, while ensuring the accuracy of AI analysis section, by building a new network structure, the system can automatically output text for the diagnosis area, These words can show the diagnosis basis of the system< br / > to this end, the research team designed a neural network system including scanner network (S-Net), diagnosis network (D-Net) and aggregator network (A-NET)These three modules play the role of image analysis, text expression, information integration and output in the system, and jointly play the role of tumor detection and cell characterization extraction< br / > the core of scanner network (S-Net) is multimodal CNN, which is a special deep neural network modelIts particularity is reflected in two aspects: on the one hand, the connection of its neurons is not fully connected, on the other hand, the weight of the connection between some neurons in the same layer is sharedIts network structure of incomplete connection and weight sharing makes it more similar to biological neural network, reduces the complexity of network model, and reduces the number of weights< br / > the diagnosis network (D-Net) acts on each ROI (region of interest, region of interest, area to be concerned selected by AI box), analyzes the pathological features and displays the feature perception network to try to explain the sketch principle of each ROI and explain what the diagnosis network sees when describing and observing, and finally converts the analysis process and results into words< br / > in short, the role of D-Net is to generate explanatory content, to tell human AI why to select these ROI and how to judge individual ROI< br / > the aggregator network (A-NET) processes the information generated by the scanner network and the diagnosis network, integrates all features, and generates the diagnosis results matching the image< br / > by scanning the pathological pictures one by one, the three modules extract the effective pixels corresponding to the database from the picture pixels and recognize them, and finally convert them into the treatable text data, and then make the system establish the direct connection between the text and the image < br / > when the data format of the network is transformed, NLP will be used to generate the language description including the characteristics of the diagnosis tissue cell and nucleus, which matches the operation mode of the pathologist The expression structure generated by NLP conforms to the clinical pathology report standard Therefore, this method can be regarded as the explanation of the artificial intelligence diagnosis process < br / > pathologists play an important role in the experiment When the pathologist processes the pathological sections, the system will capture the operation process of the pathologist, such as clicking the location of the image, and combine the operation, medical language and system language, which constitutes the logical basis of the operation and analysis of the system < br / > finally, the system can clearly explain its analysis process through its text and visual output, and provide the pathologist with direct evidence (i.e second opinion) for review and visual inspection, so as to help reduce the subjective differences in clinical decision-making of pathologists < br / > what kind of samples are used in this experiment? < br / > the data of urothelial cancer sections of nearly 1000 bladder cancer patients were used in this experiment The whole data set was divided into 620 pathological sections for training, 193 pathological sections for verification and 100 pathological sections for testing < br / > morphologically, the dataset includes 102 cases of non-invasive low-grade papillary urothelial carcinoma and 811 cases of non-invasive or invasive high-grade papillary urothelial carcinoma These data have been strictly diagnosed by many pathologists, and low-quality sections have been removed < br / > in order to evaluate the effect of neural network system, 21 urogenital pathologists participated in data annotation and diagnostic performance evaluation After nearly two years of efforts, pathologists used the web-based annotation program developed by researchers to collectively clean up and manually annotate the data < br / > by comparing the test results of the system with the routine examination of pathologists, the results show that the system achieves 97% AUC score, and its performance is better than that of most pathologists < br / > in addition, when the confounding matrix was used for comparison (Figure e, f), the average accuracy of the system was 94.6%, while that of the pathologist was 84.3% < br / > in fact, the statistical results also show that the consistent diagnosis rate of pathological doctors for some types of prostate cancer is less than 50% Therefore, only from the data point of view, the AI system proposed in this paper has good performance in accuracy and consistency < br / > the study of AI interpretability < br / > as shown above, the system explores the interpretability of AI assisted diagnosis through scanner network, diagnosis network and aggregator network, and finally produces the simultaneous output of explanatory text and ROI < br / > explanatory diagram < br / > as shown in the figure above, a and b show the results of the whole tumor detection, C, D and E are the generated "feature perception attention map", which describes the diagnosis details We can see that for each slice, the system not only selects the ROI area box routinely, but also generates explanatory text for different areas after reading < br / > the text with different characteristics is distinguished by different colors, and the ROI corresponding to the description is indicated by the box with the same color, which is convenient for the pathologist to check one-to-one correspondence < br / > the system describes a certain number of observed cell characteristics and characteristic perceptual attention map, which provides a powerful explanation for the types of visual information observed by the network (Figure C-E) In fact, the attention map contains the weight of each pixel in the frame selection area to determine the importance of different pixels for a given feature observation, but the output content is not a puzzling value, but similar to the interpretation basis of the pathologist < br / > such specialized text expression strengthens the credibility of AI analysis of pathological sections When the results of human doctors and machine diagnosis are inconsistent, doctors can also more easily compare their own diagnosis opinions with those of machines, understand the reasons for the differences, and to a large extent improve the accuracy of diagnosis < br / > evaluation of system network components < br / > in terms of algorithm structure, the performance of each part of the algorithm is verified after completion < br / > first, the researchers evaluated the tumor detection recall rate of S-Net for both tumor and non tumor images (non tumor images represent areas of clipped sliding tissue that do not highlight tumors internally) S-Net achieved 94% high true positive (number of detected tumor pixels / total annotation tumor pixels) and maintained 95.3% negative recall rate at the same time < br / > secondly, the researchers used two evaluation indicators to verify the quality of the generated diagnostic description: bilingual evaluation under study (Bleu) and consensus based image description evaluation (cider) These results show that the algorithm has some advantages < br / > this experiment breaks through the key difficulty of AI pathology three types of evidence approval < br / > limited by the unexplainability of its decision-making process, "deep learning" has been rejected by clinicians who follow the evidence-based medicine guidelines, and has become the key to restrict the development of medical image artificial intelligence, especially to obtain three types of evidence approval < br / > but this experiment provides a new way for the approval of artificial intelligence: Although the artificial intelligence at this stage still does not have reasoning ability, we can modularize the reasoning steps of doctors to simulate the reasoning process In addition, the text matching process in this experiment is in accordance with WHO standard and has strict basis, which is different from many sketches generated based on multi sample deep learning Each step of the experiment can be provided with decision-making basis by AI, not just black box operation under probability cloud < br / > Professor Yang Lin is currently the CEO of Dienga This is a solid step of Dienga technology in the declaration process of three types of certificates The interpretable bank, which is widely required in the declaration of three types of CFDA certificates, provides a key core technology solution During the interview, Professor Yang Lin also summarized the shortcomings of this experiment First of all, due to the time, the sample selection test itself has a certain closeness, with the continuous breadth and depth of data collection