Develop Prediction Model Over Incident Of Knee Osteoarthritis Based On Radiomic Features Of Infrapatellar Fat Pad

Objective:Our purpose is to(1)build a deep learning(DL)-based IPFP auto-segmented model,and(2)to set up an incident of knee radiographic osteoarthritis(IOA)prediction model with radiomics features based on auto-segmentation.Material and method:Material:The subject of this study came from Osteoarthritis Initiative(OAI),an international open resource database.(1)DL-based IPFP auto-segmentation:200 knees(40 knees for K-L grade 0-4 seperately)were randomly selected from BL except those MRI with artifact,knees underwent total knee replacement surgery or knees in IOA cohort.(2)radiomics features predict IOA in one year(IOA cohort):case knees were defined as IOA by 48-month but without knee radiographic OA(kROA)at baseline(BL).Control knees were ones not developed in kROA by 48-month matched by age(±5 years),gender and radiographic status of bilateral knees with case knees.Intermediate-weighted turbo spin-echo sequence(IW TSE)MRI at one year prior to IOA(P-1)and clinical information were collected.Method:(1)DL-based IPFP auto-segmentation:Manual segmented IPFP on each slice of MRI,then all the subjects were randomly divided into training set(70%),validation set(15%)and testing setl(15%).Testing set2 was consisted of 20 knees from group IOA and control.A deep learning method 2.5D U-net was introduced for constructing autosegmentation model.The performance of DL segmentation model was evaluated according to the Dice Similarity Coefficients(DSC)of testing set.(2)radiomics features predict IOA in one year:DLsegmentation model was used to segment IPFP of group IOA and control.2 groups were randomly divided into training set(67%)and the validation set(33%).Extracted the radiomics features of IPFP,then logistic regression was used to pick out the most significant features.Using the selected radiomic features of training set to build IOA prediction model with or without clinical characteristics(sex,age,BMI,knee injury history,knee surgery history).The performance was compared with model only used clinical characteristics,IPFP semiquantitative score(Hoffa synovitis score)or combined this two.Results:After eliminated the cases without integrated radiographic image,clinical characteristics or cannot be processed:(1)DL-based IPFP auto-segmentation:196 knees(training/validation/testing 1:136/30/30)were finally used for building DL model.DSCs of training set,testing setl and testing set2 were 0.967±0.013,0.898±0.022 and 0.900±0.019,separately.(2)radiomics features predict IOA in one year:605 knees(ncase=303,ncontrol=302)were used for built prediction models.Radiomic features with clinical characteristics model won the best performance with AUC 0.702.Radiomic features model without clinical characteristics model has the same prediction performance(AUC 0.700).The models using radiomic features as prediction factors outperform the models build up with clinical characteristics or Hoffa synovitis score only or the combination of the two(AUC 0.591,0.598 or 0.651,separately)Conclusion:The DL segementation model performed well in IPFP segmentation.The autosegmented IPFP radiomics features predict IOA in one year appropriately.Therefore,IPFP can be regarded as an independent prediction biomarker of IOA in one year.