Study On The Construction And Application Of Deep Neural Network-based Imaging Aided Diagnosis Model Of DDH

Objective:Developmental dysplasia of the hip(DDH)is a common orthopaedic disease.It is important to diagnose and treat this disease early.For a better prognosis,early diagnosis and treatment of DDH is important.Although various screening and diagnostic methods such as ultrasound and x-ray are available,it is often asymptomatic or mild in the early stages,and there are no significant changes in the imaging findings.There are also regional differences in the level of medical care,especially at the primary level.There is a lack of specialized orthopedic surgeons.A large number of patients are still being missed.Without proper treatment,most of the patients' symptoms gradually worsen in their 20 s and 40 s and develop into hip arthritis that requires joint replacement.There is an urgent clinical need for a simple and cost-effective scientific tool to screen a large number of orthotopic pelvic radiographs to assist in the early diagnosis of DDH and initial assessment of its severity.There is also such a clinical need for internal and external knee arthroplasty.This paper presents modified Mask R-CNN model and segmentation model for model training and testing to construct artifacts for measuring the sharp,CEA and HKA angles to aid in the diagnosis of DDH and knee valgus.Screening of high-volume orthotopic pelvic X-rays and full-length lower limb films is performed to reduce the missed misdiagnosis of DDH and knee valgus.Methods:12427 pelvic orthotopic X-rays and 469 lower extremity full-length X-rays are collected from the PACS system of the Second Hospital of Jilin University,of which 101,104,100,and 100 X-rays are randomly selected for validation testing in four trials,and the remaining X-rays are labeled,pre-processed,and entered into the modified Mask R-CNN model and segmentation model for model training and testing,respectively.The four new artificial intelligence models are constructed.The newly constructed AI models are validated by measuring sharp angle,CE angle and HKA angle.The diagnostic accuracy of the AI model for diagnosing DDH based on angular measurement was compared and analyzed with the diagnostic accuracy of 100 pelvic orthotopic films by 30 orthopaedic surgeons.Results:1.Using the labeled data of 11,473 pelvic orthotopic slices images,we entered the modified deep neural network model training to build an AI model for automatic measurement of acetabular shapr angle.The left and right sides of its AI model measurement shapr angle are 40.067 ± 4.087° and 40.653 ± 4.214°,and the average values of the left and right sides measured by three doctors are 39.353 ± 6.738° and 39.821 ± 6.986°.There is no significant difference between the AI model group and the doctors' group measurements,which validates the AI model's ability to predict key feature points and measure the sharp accuracy.The AI model takes 120 seconds to measure the sharp angle of 101 orthotopic pelvic slices,and the physician takes an average of 150 minutes.The AI model shows an absolute advantage in terms of measurement efficiency,and the AI model and the doctor's measurement of the sharp angle are used to evaluate the agreement between the pelvic acetabulum results and the Confirmed diagnosis test results.As shown by the kappa test results,the AI model basically agrees with the Confirmed diagnosis results(P < 0.05).2.The fine-tuned Mask R-CNN is trained to construct an artificial intelligence model for automatic measurement of acetabular CEA using data from 13,228 labeled pelvic orthotopic slices images.The AI measures the right and left lateral mean values of CE angle as 29.46 ± 6.98° and 27.92 ± 6.56°;the three physicians measure the right and left lateral mean values of CE angle as 29.85 ± 6.92° and 27.75 ± 6.45°.The paired-sample t-test analysis shows no statistically significant difference between the left and right side AI measurements and the mean physician measurements(P>0.05).The Kendall W consistency test is chosen to obtain the left Kendall W=0.994,P<0.001;right Kendall W=0.995,P<0.001,and the AI model shows a high degree of agreement with the physician measurements,verifying the accuracy of the model in identifying key points and measuring the acetabular CE angle.3.Based on a modified deep neural network model,an artificial intelligence(AI)model for the automatic measurement of acetabular sharp angle and CE angle to assist in the diagnosis of DDH has been constructed by deep learning of the data labeled on 12,225 pelvic orthotopic slices.When comparing this AI model with the one used by 30 doctors to diagnose DDH through 100 pelvic orthotopic slices,the AI model takes 134 seconds and scores 93 points;the chief physician group takes an average of 576.2 seconds and scores 83.4 points;the associate chief physician group takes an average of 916.1 seconds and scores 66.4 points;the attending physician group takes an average of 557.0 seconds and scores 50.8 points;all the doctors take an average of 576.2 seconds and scores 83.4 points;the chief physician group takes an average of 557.0 seconds and scores 50.8 points.Doctors take an average time of 683.1 seconds and have an average score of 66.9.4.Based on the segmentation algorithm,an artificial intelligence model for automatic measurement of HKA angle is constructed by deep learning of 738 full-length lower limb slices labeling data.The segmentation performance of the three networks of the model is assessed.The Dice,Recall and Precision values for each deep neural network are 84.13%,83.21% and 86.42% for the femur;93.45%,91.17% and 96.16% for the knee;and 81.79%,76.80% and 88.71% for the ankle,respectively.The mean physician-measured mean angle(agreed true value)versus standard deviation is169.32°±9.88°,while the mean predicted value versus standard deviation is169.81°±9.61°.The ICC agreement between the agreed true value and the predicted value is high.The accuracy of the model for dividing the femoral head,knee,hip and calculating the center point for measuring the HKA angle is verified.Conclusion:In this study,four new and effective medical AI models are successfully constructed by modifying the Mask R-CNN algorithm and segmentation algorithm,inputting the labeled large number of pelvic orthotopic X-rays and full-length lower limb image data into the models for training and testing.The new artificial intelligence models can accurately detect key points such as the lower edge of the teardrop on the X-ray film,the center of the femoral head,the outer edge of the acetabulum,the center of the knee joint,and the center of the ankle joint,and automatically plot the measurement of the sharp angle,CE angle,and HKA angle according to the identified key points.The clinical diagnosis codes corresponding to different degrees are written into the model,and the model gives diagnostic opinions based on the measured angles according to the diagnostic criteria.This provides a new intelligent measurement tool for orthopedic imaging measurement and evaluation.It also provides a new and reliable method of diagnostic screening for DDH and knee valgus for primary and less diagnostically experienced physicians.As well as validating the great advantages of Mask R-CNN model and segmentation algorithm in orthopedic imaging measurement through this study,it lays a research foundation for orthopedic imaging measurement.