Research On Artificial Intelligence Algorithm And Its Application In Medical Diagnosis

Medical information is different from the natural environment.Medical information has structure features or functional features,and and it is easy to appear that the region of interest(Ro I)is small,the similarity with background is high,and the fine texture structure needs to be considered.And different diseases demonstrate different characteristics and represent heterogeneous.Therefore,ensuring machines to support decision of medical diagnosis is one of the most elusive and long-standing challenges.On the one hand,we investigate algorithm of artificial intelligence to improve reading effectiveness on large-scale medical data and precise medical diagnosis(black-box),and provide advanced novel automatic tools and ideas for the research of medical data analysis;On the one hand,we tackle the interpretable margin of artificial intelligence and provide the interpretability of the algorithm(white-box)under complex medical phenomena to enhance its clinical utility.The main research contents and important innovations summarized as follows:(1)A crucial information localization and segmentation algorithm for intelligent diagnosis of chronic middle ear diseases: Chronic middle ear diseases play important roles in daily otorhinolaryngology practice due to their high incidence.In today’s hospitals,otolaryngologists usually obtain a comprehensive understanding of the structure of the ME through endoscopy and CT images and make a treatment plan based on the diagnosis.The main purpose of this study was to develop a an intelligent diagnosis algorithm for different chronic middle ear diseases based on crucial information localization and segmentation.First,to solve the data imbalance problem caused by a small number of cholesteatoma cases,image inversion was employed for data augmentation.Then,the crucial information of interest on the brain CT scans are extracted after located.Finally,We used transfer learning to simplify the training of similarity classification networks on both sides.This study provides a new automatic detection direction for the discrimination of middle ear diseases,and provides doctors with an unbiased diagnosis reference before diagnosis.The proposed scheme shows the best performance for identification of chronic middle ear diseases in CT scan images.(2)A long-short temporal-spatial sequence feature tracing algorithm for intelligent detection of esophageal disorders: Esophageal high-resolution manometry is widely performed to evaluate the representation of manometric features for diagnosing esophageal disfunction,but there has been few research work on artificial intelligence.In this study,a deep tracing model based long-short temporal-spatial sequence is proposed to learn location,pressure,time,and trajectory-related features for anomaly detection.This work investigates the trajectories of esophageal functional contractility,performs a functional comprehensive analysis,and extends it to the detection of esophageal disorder.The findings show that for each patient high-resolution esophageal manometries,the proposed method takes only 1.03 seconds to identify esophageal disorders,which is faster than clinical experts and can be a very useful diagnosis tool,can be used inexpensively and quickly during esophageal manometry testing.(3)A class-confidence map representation algorithm for intelligent identifying of COVID-19 and the interpretable perception of infected tissue: The COVID-19 pandemic continues to threaten the health and lives of billions of people.This work develops and tests an efficient and accurate artificial intelligence algorithm to help radiologists automatically identify COVID-19 and intelligently mark COVID-19 infected tissue.Compared with the radiologists,the accuracy of COVID-19 discrimination yielded98.71%,while the accuracy of localization was 93.03%.The most important contribution of this work is to provide a class-confidence map to represent distribution of crucial class feature.Class-confidence map representation provided highly comprehension and demonstrated the feasibility to hoc-post interpret the outcome of black-box AI.This work not only effectively and accurately distinguishes COVID-19,but also precisely percepts infected tissue,which is highly consistent with radiologists.(4)A structure-constrained deep feature fusion algorithm for intelligent detection of middle ear diseases: Currently,the analysis of the middle ear structure by temporal bone CT has become an important clinical diagnostic method of middle ear disease,because the eroded parts of the middle ear are different.We first used a region-of-interest(ROI)network to find the middle ear in the entire CT image and segment it to a size of 100 × 100 pixels.Then,we used a structure-constrained deep feature fusion algorithm to convert different structures of the middle ear based on chronic suppurative otitis media,middle ear cholesteatoma and normal patches.To fuse structure information,we introduced a graph isomorphism model,that implements a feature vector from neighborhoods and the coordinate distance between vertices,thus adjusting the neighborhood weight in the information cluster by adding a control gate unit.The experimental outcomes demonstrated an accuracy of 96.36% in the classification of middle ear diseases.The proposed algorithm discriminates quickly and efficiently by analyzing the structural similarity between the middle ear diseases,reducing the workload of physicians and improving their treat quality.(5)An implicit propagation prediction algorithm of vigor for recognizing esophageal contraction patterns: This study proposes a efficient and interpretable model for propagation prediction of esophageal vigor.We designed an efficient graph to incorporate contractile vigor propagation that considers the contraction and pressure propagation between vigor in highresolution manometry.In this research,we used convolutional neural network to implement the extraction and graph construction of contractile vigor in high-resolution esophageal manometry data.Using an attention graph convolutional network,the edges in contractile vigor propagation can automatically learn the contraction patterns trends in time series and propagation through the attention units.The attention mechanism layer leverages the short-term trend to improve the prediction accuracy.Based link prediction,the propagation of the contractile vigor graph are end-to-end automatically optimized.The proposed algorithm can not only extract the contractile vigor in a single swallowing scene,but also generate a contractile vigor propagation graph for further analysis and recognition of manometries to perform diagnose.