| Femoral trochlear dysplasia(FTD)refers to the abnormal depth or geometry of the femoral trochlear groove,which is an important risk factor for patellar instability.The structural abnormality can accelerate the wear of patellofemoral joint surface and induce patellofemoral arthritis.Standard knee lateral radiographs play an important guiding role in diagnosing FTD,which are not common in clinical practice.MRI can clearly display the anatomy of articular cartilage,ligaments,muscles and the morphology of femoral trochlea,which has certain advantages in diagnosing FTD.According to the literature review,the diagnosis of FTD involves a variety of indicators,among which the femoral trochlear groove angle,lateral trochlear inclination,depth of trochlear groove,and trochlear facet asymmetry perform well in distinguishing normal and abnormal femoral trochlea.Classification also includes Dejour type,OBC type and arthroscopic morphological type,among which Dejour type is the most widely used classification system of FTD.Although there are many studies on the diagnosis and classification of FTD,there are still some shortcomings.First,there are many indicators.Relying only on manual measurement increases the work burden of orthopedic doctors,and time-consuming operations will increase the incidence of misdiagnosis.Second,the classification of FTD depends more on the clinical experience of doctors.The diagnosis advantage of senior doctors is far more than that of junior doctors.The misdiagnosis of junior doctors often occur.In recent years,artificial intelligence(AI)in the field of orthopedics has been rapidly developed.Application areas mainly include: detection,classification and image segmentation.These areas can be combined with each other.The characteristics of high efficiency and stability of AI model greatly shorten the diagnosis timeliness and ensure the diagnosis objectivity.These advantages can completely improve the current deficiencies in the diagnosis and classification of FTD.But no similar studies have been reported so far.Based on the above reasons,the following three parts of research are carried out for assisted diagnosis and classification of FTD by using AI technology.Part 1: Qualitative Study on Assisted Diagnosis and Classification of Femoral Trochlear Dysplasia by Using AI Technology Purpose:To explore the views of specialists on assisted diagnosis and classification of FTD by using AI technology.To provide professional advice on model construction,and to understand the potential benefits and limitations of using AI to assist in diagnosing FTD.Method:In this study,the descriptive research method and maximum difference sampling strategy were adopted.Fourteen experts of orthopedics and radiology from different hospitals in Jilin Province were selected as research objects from September 2020 to January 2021.Data were collected in the form of semi-structured and one-to-one interview.Qualitative analysis software of NVivo 2020 was used to organize and analyze the interview data.Results:Through in-depth interviews and comprehensive analysis of 14 orthopedics and radiology experts from different hospitals,four core themes were obtained: "Diagnosis cognition of femoral trochlear dysplasia","Adience and avoidance conflict in AI diagnosis","Attitude and suggestions on AI diagnosis model" and "Human and technology in medicine".Conclusion:Most orthopedics and radiology experts support the use of AI in diagnosing FTD.The AI model will provide help for junior doctors,and have potential value in improving the diagnosis speed of FTD,ensuring the diagnosis accuracy and coordinating medical resources.However,in the construction and use of AI models,attention should be paid to data quality and data security.This study explores the diagnosis and classification indicators commonly used by specialists in clinical work,and provides a clinical practice basis for constructing AI models in future.Part 2: Study on Constructing Assisted Diagnosis Model of Femoral Trochlear Dysplasia by Using Improved Heatmap Regression Method Purpose:To construct a diagnosis model for measuring quantitative indicators of femoral trochlea and assist orthopedics doctors to diagnose femoral trochlear dysplasia.Method:A total of 464 knee MRI cases,including FTD(n=212)and normal femoral trochlear(n=252),were collected from January 2019 to December 2020.After data processing by the senior orthopedics doctor,and reviewing by the senior radiology and orthopedics experts with work experience of more than 20 years,94 cases were randomly selected as the test set without joining the model training.The remaining images were fed into the improved AI model for training and verification.In order to achieve accurate automated location of key points,the research uses heatmap regression method.After the key points are identified,the measurement results are obtained according to the formula constructed by the diagnosis indicators.The AI model and two orthopedics doctors respectively measured lateral trochlear inclination,depth of trochlear groove and trochlear facet asymmetry,and then made the diagnosis based on the measurement results.The diagnosis results were statistically analyzed.The area under ROC curve and accuracy were used to compare the diagnosis efficiency,and the kappa value quantified the intergroup and intragroup consistency.Results:Using the measurement results of senior orthopedics experts as reference standard,the accuracy,sensitivity,specificity,positive predictive value,negative predictive value and AUC value of AI model in diagnosing FTD were 0.88,0.79,0.96,0.94,0.84,0.88,all of which were better than that of intermediate doctors and junior doctors.However,the diagnosis time is much lower than that of junior and intermediate doctors.In terms of consistency with reference standard,kappa value of AI model was 0.76,which was highly consistent and superior to intermediate doctors and junior doctors.Conclusion:The AI model was constructed by using the improved heatmap regression method.By introducing the appropriate attention mechanism in the image,the highest point of the medial and lateral femoral condylar,the lowest point of the trochlear groove,and the lowest point of the medial and lateral femoral posterior condylar could be accurately located.By measuring the depth of trochlear groove and other quantitative indicators to diagnose FTD,the diagnosis efficacy is better than that of intermediate and junior orthopedics doctors.Part 3: Study on Constructing Automated Classification Model of Femoral Trochlear Dysplasia by Using Improved Res Net Network Purpose:An automated classification model of FTD was established based on Res Net network.It can assist clinicians to correctly classify FTD as mild or severe,so the clinicians could make the correct treatment plan.Method:From January 2016 to October 2022,955 cases of MRI diagnosed as FTD were collected from our hospital.Radiology and orthopedics experts with work experience of more than 20 years classified all the samples into mild or severe based on Dejour classification,including441 cases of mild and 514 cases of severe.The knee MRI samples were processed by a senior radiology doctor with 13 years of experience.192 cases were randomly selected as the test set without joining the model training.The remaining 763 cases were input into the AI model for training and verification.On the basis of the second part,convolutional neural network and recurrent neural network technology were added to construct three deep learning models.The purpose was to complete the automated selection of the first craniocaudal image where the trochlear cartilage can be reliably visualized,the identification of key regions containing the femoral trochlear groove and the classification of femoral trochlear dysplasia.Finally,AI model and three doctors respectively classified the samples of test set into mild or severe.The area under ROC curve and accuracy were used to compare the classification efficiency.Cohen’s kappa and Fleiss kappa tests were used for statistical analysis,and kappa value quantified the intergroup and intragroup consistency.Results:The classification results of senior radiology and orthopedics doctors were used as the reference standard.Based on the the first craniocaudal image with trochlear cartilage,the sensitivity,specificity,positive predictive value,negative predictive value,accuracy and area under ROC curve of the semi-automated classification model to classify FTD were 0.93,0.93,0.92,0.94,0.93,0.93.Based on all axial MRI images,the sensitivity,specificity,positive predictive value,negative predictive value and accuracy of the automated classification model to classify FTD were0.78,0.86,0.83,0.82,and 0.82.The area under ROC curve were 0.82 for the automated classification model,0.85 for senior doctors,0.78 for intermediate doctors,and 0.69 for junior doctors.The classification efficiency of the automated classification model is comparable to that of senior doctors and intermediate doctors,but better than that of junior doctors.In terms of consistency with reference standard,the kappa value of the automated classification model was 0.64,slightly lower than that of the senior doctors(0.70),and better than that of the intermediate doctors(0.55)and junior doctors(0.38).The intragroup consistency of the automated classification model was better than that of all doctors.Conclusion:The classification model constructed by convolutional neural network,key point detection,attention mechanisms and recurrent neural network technology can automatically classify FTD into mild or severe based on knee MRI.The classification efficiency is slightly lower than that of senior doctors,but better than that of intermediate doctors and junior doctors.There is no similar research report so far.The AI model provides a reliable classification tool for junior orthopedics doctors with little experience,especially for orthopedics doctors in underdeveloped areas.The research further proves the application value of deep learning technology in the field of orthopedics. |
PDF Full Text Request