| ObjectiveOsteoarthritis(OA)is one of the main causes of disability in elderly people and has high socioeconomic costs.Due to its slow progression and irreversible joint structural changes,there is currently no disease-modifying therapy available for OA.The meniscus and articular cartilage of the femorotibial joint are crucial for protecting the femorotibial joint space.Previous studies have shown that MRI-based artificial evaluation indices of meniscus and femorotibial cartilage have predictive value for the incidence of knee osteoarthritis(KOA),but there are currently no reports on the establishment of a radiographic KOA prediction model using combined meniscus and femorotibial cartilage imaging radiomics.Given that OA is a disease that affects the entire joint,this study aims to use a convolutional neural network to automatically segment the MRI images of femoral cartilage,tibial plateau cartilage,lateral meniscus,and medial meniscus of volunteers with pre-radiographic KOA,and to establish a radiographic KOA and total knee replacement(TKR)prediction model by combining radiomics from these four anatomical regions through neural networks.This will provide a novel non-invasive tool for the clinical discovery of radiographic KOA and assist in the early detection and personalized intervention of radiographic KOA patients.MethodsBased on MRI data from the Osteoarthritis Initiative(OAI)multicenter cohort study,participants who had a high risk of developing radiographic knee osteoarthritis(KOA)but had no evidence of radiographic KOA at baseline(defined as a knee joint space narrowing score of<2 on a standing X-ray image,indicating the absence of radiographic KOA),and who were over 45 years of age,were selected and named the Incidence of OA cohort(Inc.OA)for the outcome of radiographic KOA.Participants who had completed a 4-year follow-up,had no history of TKR at baseline,but had undergone TKR within 4 years of follow-up were selected and named the Progression of OA cohort(Pro.OA)for the outcome of total knee replacement.For the outcome of radiographic KOA,343 knee joints were selected as the case group based on the occurrence of radiographic KOA within 4 years of follow-up,and 343 knee joints were selected as the control group,matched for age(within 5 years),baseline KLG score,gender,and contralateral knee joint status.For the outcome of TKR,219 knee joints were selected as the TKR group based on the occurrence of TKR within 4 years of follow-up,and 223 knee joints were selected as control group 1,matched for age(within 5 years),baseline KLG score,gender,and contralateral knee joint status.Using baseline sagittal 3D double echo steady-state with water excitation(3D-DESS-WE)MRI,image grayscale mapping was used to display the grayscale values in the regions of interest(ROI)in the 3D images using heat maps,allowing us to visually identify the differences between the Case group and Control group,as well as the differences between the meniscus,femoral cartilage,and tibial cartilage in the TKR group and Control group.The knee joints were then randomly divided into training and validation groups(8/2 ratio)using 4-fold cross-validation(repeated 2000 times)for the Inc.OA and Pro.OA cohorts.A convolutional neural network was used for automatic image segmentation(segmented into femoral cartilage,tibial cartilage,lateral meniscus,and medial meniscus),and feature extraction was performed based on the segmentation using the open-source pyradiomics code package.The following models were established using neural networks:femoral cartilage radiomics model(FC-RM),tibial cartilage radiomics model(TC-RM),lateral meniscus radiomics model(LM-RM),medial meniscus radiomics model(MM-RM),multi-structure radiomics model(MS-RM),and multi-structure radiomics plus clinical variable model(MS+C-RM).The performance of the models was evaluated using the area under the curve,sensitivity,specificity,accuracy,precision,F1 score,Matthews correlation coefficient,and kappa coefficient on the training and validation sets for each of the group models.ResultsIn the POMA_inc.OA queue,a total of 549 knees were included in the training set(275 cases and 274 controls),and 137 knees were included in the validation set(68 cases and 69 controls).In the validation queue,the AUC(95%CI)for FC-RM,TC-RM,LM-RM,and MM-RM were 0.793(95%CI:0.707-0.858),0.725(95%CI:0.634-0.801),0.816(95%CI:0.736-0.876),and 0.803(95%CI:0.720-0.865),respectively.In addition,the kappa coefficients for FC-RM,TC-RM,LM-RM,and MM-RM ranged from 0.299 to 0.460.In the training set,MS-RM showed good accuracy in predicting the incidence of radiographic knee osteoarthritis,with an AUC(95%CI)of 0.955(95%CI:0.938-0.968).In the validation set,the AUC for MS-RM was 0.931(95%CI:0.876-0.963),with a sensitivity of 0.844(95%CI:0.839-0.849),specificity of 0.856(95%CI:0.852-0.860),precision of 0.857(95%CI:0.853-0.860),accuracy of 0.850(95%CI:0.847-0.853),F1 score of 0.850(95%CI:0.847-0.853),MCC of 0.701(95%CI:0.694-0.707),and kappa coefficient of 0.694.In the validation set,MS-RM outperformed the single-structure radiomics models(FC-RM,TC-RM,LM-RM,and MM-RM)in predicting knee osteoarthritis,with an increase in AUC by 14.2-28.4%,sensitivity by 15.0-44.3%,specificity by 16.5-33.3%,precision by 18.9-34.4%,accuracy by 21.1-36.4%,F1 score by 19.9-39.3%,MCC by 72.7-100.3%,and kappa coefficient by 50.9-131.3%;all p-values were less than 0.001.In the POMA_pro.OA cohort,a total of 355 knee joints were included in the training set(180 cases and 175 controls),and 87 knee joints were included in the validation set(39 cases and 48 controls).In the validation set,the AUC(95%CI)for FC-RM,TC-RM,LM-RM,and MM-RM were 0.718(95%CI:0.600-0.812)、0.829(95%CI:0.728-0.898)、0.727(95%CI:0.608-0.820),and 0.673(95%CI:0.547-0.778),respectively.In addition,the kappa coefficients for FC-RM,TC-RM,LM-RM,and MM-RM ranged from 0.189 to 0.516.In the training set,MS-RM showed good accuracy in predicting the incidence of radiographic knee osteoarthritis,with an AUC(95%CI)of 0.933(95%CI:0.904-0.954).In the validation set,the AUC for MS-RM was 0.931(95%CI:0.864-0.966),with a sensitivity of 0.832(95%CI:0.826-0.837),specificity of 0.805(95%CI:0.800-0.810),precision of 0.777(95%CI:0.773-0.781),accuracy of 0.817(95%CI:0.813-0.820),F1 score of 0.803(95%CI:0.799-0.806),MCC of 0.635(95%CI:0.628-0.641),and kappa coefficient of 0.676.In the validation set,MS-RM outperformed the single-structure radiomics models(FC-RM,TC-RM,LM-RM,and MM-RM)in predicting knee osteoarthritis,with an increase in AUC by 12.3-38.5%,sensitivity by 12.6-36.2%,specificity by 13.1-36.7%,precision by 14.8-40.3%,accuracy by 12.8-36.4%,F1 score by 13.8-39.5%,MCC by 41.2-201.9%,and kappa coefficient by 31.0-257.7%;all p-values were less than 0.001.ConclusionUsing radiomics features extracted from the MRI of meniscus,femoral and tibial plateau cartilage,we developed a neural network model that accurately predicts the incidence of radiographic knee osteoarthritis(KOA)and TKR.This model can help identify meniscus-femoral-tibial plateau cartilage changes in pre-radiographic OA patients and radiomics changes in patients before knee replacement,providing insights for early clinical intervention and personalized treatment.However,the clinical utility of the model needs to be further validated in future studies. |
PDF Full Text Request