Predicting The Prognosis Of Osteonecrosis Of The Femoral Head After Huo Xue Tong Luo Method Treatment Based On Magnetic Resonance Imaging And Machine Learning

ObjectiveIn hip preservation treatment for osteonecrosis of the femoral head(ONFH),precise assessment of the risk of femoral head collapse progression is important for the selection of treatment strategy.Although the efficacy of hip preservation therapy with Huo Xue Tong Luo method is clear,it is still necessary to develop a personalized treatment plan according to the risk of collapse progression.Improvement of femoral head collapse progression prediction is still needed.This is mainly due to the lack of three-dimensional assessment of the anterolateral integrity of the femoral head and the inability to quantify the changes in bone structure within the necrotic lesion.Therefore,the main objectives of this study were:(1)to develop a method for quantitative assessment of necrotic lesions on three-dimensional images and to investigate the value of this method in assessing the risk of collapse progression;(2)to analyze the heterogeneity of bone structure of necrotic lesions on magnetic resonance imaging(MRI)with radiomics,and use machine learning to integrate features reflecting the location range of necrotic lesions and the bone structure of necrosis lesion to establish a relatively accurate and dynamic model for femoral head collapse progression prediction.MethodsStudy Ⅰ:The study included 206 hips with ONFH.Coronal views of T1 sequence MRI were acquired,of which 190 cases were randomly divided into training set and 16 cases into test set.The images were cropped and necrotic lesions were manually segmented,which was used as the ground truth for model training.Training of the segmentation model was performed using the nnU-Net algorithm.Five-fold cross-validation method was performed on the training set,and the final validation was performed on the test set using the ensemble model.The performance of the model was evaluated using the Dice similarity coefficient(DSC)and 95%Hausdorff distance(95%HD).Study Ⅱ:A retrospective case series analysis was performed on 131 hips with ONFH from January 2017 to December 2020.MRI of the affected hips were acquired and the necrotic lesions were segmented and visualized using the automatic segmentation model trained in Study Ⅰ,and the extent of necrotic lesions was analyzed layer by layer.The ratio of the layer count of necrotic lesions to the layer count of the femoral head was calculated,which was defined as necrosis ratio(NR),and the ratio of the layer count of large necrotic lesion to the layer count of the femoral head was calculated,which was defined as large lesion ratio(LLR).The differences in NR and LLR were compared between the progressive and non-progressive groups,and subgroup analysis was performed by stratifying the cases according to the Japanese Investigation Committee(JIC)classification.Operating characteristic curves(ROC)analysis was performed to investigate the correlation between NR and LLR and progression of collapse.Finally,affected hips were grouped according to scanning parameters to investigate the influence of scanning parameters on NR and LLR.Study Ⅲ:210 affected hips with ONFH from January 2017 to December 2020 were retrospectively included,and 159 affected hips from the First Affiliated Hospital of Guangzhou University of Chinese Medicine were split into training set,the rest 51 affected hips from the Third Affiliated Hospital of Guangzhou University of Chinese Medicine were split into test set.The NR and LLR of each affected hip were calculated using the method established in Study Ⅱ.The radiomics features of the necrotic lesion region on the coronal view of T1 images were extracted and merged with the NR and LLR.The combined MRI features were selected based on intra-group correlation coefficient,Spearman correlation coefficient,and the ten-fold cross-validated Least Absolute Shrinkage and Selection Operator.Based on the filtered MRI image features,Support Vector Machine(SVM).K Nearest Neighbors(KNN),Random Forest(RF),eXtreme Gradient Boosting,(XGBoost),Light Gradient Boosting Machine(LightGBM),and Adaptive Boosting(AdaBoost)were used to build prediction models of collapse progression and were externally validated with the test set.The performance of the prediction models was compared by plotting ROC curves and calculating the areas under the curve(AUC).The F1 scores were determined based on the precision and recall of the models and were also used to evaluate the performance of the models.The accuracy of each model’s prediction was also calculated.Finally,Decision Curve Analysis(DCA)was performed to assess the efficacy of the models in clinical application.Study Ⅳ:Fifteen cases with ONFH who received Huo Xue Tong Luo method hip preservation treatment and regular MRI follow-up from January to December 2020 in the Third Affiliated Hospital of Guangzhou University of Chinese Medicine were selected.The NR and LLR of the affected hips before and after treatment were calculated,and the radiomics features selected in Study Ⅲ were extracted.In addition,the volume of necrotic lesions before and after treatment was calculated.Based on the results of Study Ⅲ,the model with the best performance was selected to predict the risk of collapse progression using MRI features both before and after treatment.The predicted probability of the model was used as the measurement of the risk of collapse progression.The changes in MRI features and necrotic lesion volumes before and after treatment were compared.Preliminary analysis of the relationship between the changes in the model-predicted progression risk before and after treatment and the occurrence of collapse progression was performed.ResultsStudy Ⅰ:Five-fold cross-validation was performed on the training set with 100 epochs per fold,and in the five-fold cross-validation,the mean DSC with the best performance in the validation set was 0.8348 and the mean 95%HD was 4.0259:the mean DSC with the best performance was 0.8940 and the mean 95%HD was 2.4863.The performance of the ensemble model from the five models trained by the five-fold cross-validation show that the mean DSC in the test set was 0.8476 and the mean 95%HD was 3.7936,indicating that the model can achieve a relatively accurate segmentation.Study Ⅱ:Of the 131 included hips,collapse progression occurred in a total of 63 hips.The included hips were grouped by whether collapse progression was occurred.Both NR and LLR were significantly higher in the collapse progression group(P<0.001,both).Subgroup analysis showed that both NR(P=0.009 and P=0.003,respectively)and LLR(P<0.001 and P=0.002,respectively)were significantly higher in the collapse progression group in JIC type C1 and type C2 ONFH.Logistic analysis showed that both NR(OR,1.45[95%CI.1.04-2.04];P=0.031)and LLR(OR,1.46[95%CI,1.22-1.75];P<0.001)were independent risk factors of collapse progression,and both showed positive correlation with collapse progression.ROC curve analysis showed that the AUC of NR and LLR were 0.74 and 0.84,and Delong test indicated that the predictive value of LLR for progression of collapse was significantly higher than that of NR(P=0.007).Two multi-variables regression models were constructed by including NR,LLR,and ARCO staging,or further adj usted with age and sex.The AUC of both models were 0.87,and the differences were not statistically significant compared to LLR(P=0.649 and P=0.559,respectively),indicating that LLR had similar predictive value of collapse progression as the multi-variables regression model.According to the Youden index,the best threshold of LLR for predicting collapse progression was 40.00%,with a sensitivity of 88.89%and a specificity of 69.12%.When grouped by MRI scanning parameters,there was a significant difference in the layer count of femoral head and the layer count of necrotic lesion between different parameter groups(all P<0.001),but no significant difference in the layer count of large lesion(P=0.240).The differences in JIC classification,NR and LLR between different parameter groups were not statistically significant(P=0.163,P=0.320,P=0.917),indicating that the distribution of ONFH types was relatively consistent,and NR and LLR may not be affected by the differences in scanning parameters.Study Ⅲ:After screening,seven MRI features were finally selected,which were LLR,NR.first-order variance,first-order kurtosis,gray level dependence matrix(GLDM)gray level non-uniformity,GLDM gray level variance,gray level run length matrix(GLRLM)long run length low gray level emphasis.The models built by SVM,KNN,LightGBM.and AdaBoost showed good prediction performance,except for the models built by RF and XGBoost,which showed overfitting.The model built by LightGBM algorithm showed better performance compared to others,with an AUC of 0.911,an F1 score of 0.902,and an accuracy of 82.39%on the training set,and an AUC of 0.902,F1 score of 0.800,and accuracy of 78.43%on the test set.DCA curve analysis showed that when the threshold probability is between 0.15 and 0.75,the LightGBM prediction model yields net benefit.In the LightGBM model,the rank of feature importance in a descending order was as follows:GLDM gray level non-uniformity,LLR,GLDM gray level variance,first-order variance,and NR,and GLRLM long-travel low gray emphasis and first-order kurtosis were the least important.Study Ⅳ:After 6 months of Huo Xue Tong Luo method hip preservation treatment,GLDM gray level non-uniformity(P=0.023)and necrotic lesion volume(P=0.001)were significantly lower after treatment,but LLR(P=0.167),NR(P=0.317),first-order variance(P=0.176),first-order kurtosis(P=0.061),and GLDM gray level variance(P=0.491),GLRLM long run length low gray level emphasis(P=0.496),and the risk of collapse progression predicted by the LightGBM model were not statistically different before and after treatment.When grouped by whether collapse progression occurred,it was found that the risk of progression was relatively higher in hips with collapse progression both before and after treatment.Further analysis revealed that hips with risks maintaining or increasing to greater than 50%after treatment experienced collapse progression,while hips with risks maintaining or descending to less than 50%did not experience collapse progression.Conclusions(1)Calculation of NR and LLR by analyzing each slice of the femoral head on MRI coronal images is a reliable tool to quantify the location and extent of necrotic lesions in ONFH in three dimensions,and both are independent risk factors for femoral head collapse progression.LLR can effectively assess the integrity of the anterolateral column of the femoral head and is valuable for predicting femoral head collapse progression.(2)Using LightGBM algorithm to integrate indicators reflecting the location and extent of necrotic lesions and radiomic features reflecting the heterogeneity of bone structure of necrotic lesions can establish a prediction model for the progression of collapse,and provide a dynamic and quantitative assessment of ONFH.The assessment of the risk of collapse progression of the femoral head based on this model has the opportunity to become one of the imaging indicators to evaluate the efficacy of hip preservation treatment with Chinese medicine.