| Objective To explore the predictive value of the machine learning models which constructed by magnetic resonance T2 WI combined with ADC imaging features and clinical indicators in bone metastatic deposits in prostate cancer patients.Methods A retrospective analysis was done comprising of 152 patients with prostate cancer,which was confirmed by histopathology and underwent multi-parameter magnetic resonance examination.the age ranged from 45 to 87 years old,with an average of(72.4 ±8.51)years,including 80 cases with bone metastasis and 72 cases without bone metastasis.It is divided into training set and test set according to the proportion of 7:3.Based on magnetic resonance ADC and small-field high-resolution T2 WI cross-sectional images,the largest section of the prostate tumor was delineated as the region of interest,and the imaging features were extracted.The imaging score(Rad-score)of each PCa patient was calculated.Eight different machine learning prediction models are constructed by selecting two kinds of preprocessing components,two kinds of feature screening components and two kinds of AI operation components(support vector machine and logical regression algorithm).After cross-verification,the diagnostic efficiency of the four machine learning models with good stability is evaluated by using the receiver operating characteristic curve(ROC),the maximum area under the curve(AUC),accuracy,average accuracy and so on.At the same time,the clinical features,including prostate specific antigen(PSA),Gleason score,serum alkaline phosphatase and clinical TNM stage,were analyzed by univariate correlation and multivariate Logistic regression analysis to screen the risk factors and independent predictors of bone metastasis of prostate cancer.A comprehensive prediction model combined with imaging and clinical features was constructed to evaluate the prediction efficiency of imaging model,clinical risk factor model and joint prediction model.Intra-and-inter-classs correlation coefficient(ICC)or Kappa test were used to evaluate the reproducibility when drawing it again.Results Two radiologists(reader1 and reader2)of experience in prostate cancer segmented the ROI from the index cancer for assessing the ICC reproducibility,the ICC was 0.865(0.770～0.923);and the reader1 repeat the same procedure later,the ICC was 0.925(0.891～0.948).radiomics signature consisting of 15 selected features was significantly correlated with bone metastasis of prostate cancer(P < 0.01).It includes three Sfov T2 WI taxonomic features and 12 ADC sequence genomic features.Six of them belong to Shape2D(two-dimensional shape)features.Based on the standardization of the optimal feature screening algorithm-LR model AUC is 0.85 (0.75～ 0.94),AP 0.80,AP 0.78,and the test set AUC is 0.77(0.55～ 0.69),AP0.69,ACC 0.78.Each numerical index fluctuates less and more stably in the training set and test,so it is the final optimized imaging model.With the addition of clinical independent risk factors,the effectiveness of the combined predictive model in the differential diagnosis of bone metastasis of prostate cancer was further improved,the training set AUC was 0.88(0.820～0.941),and the test set AUC was 0.81 (0.692～0.922).The results of univariate and multivariate analysis of clinical features showed that Gleason score,clinical T stage,PSA and ALP were correlated with bone metastasis.Multivariate regression analysis showed that PSA and Gleason scores were independent risk factors for bone metastasis(OR >1),and Rad-score was a protective factor for bone metastasis(OR <1).Conclusion The machine learning model based on the imaging features of small-field high-resolution T2 WI and ADC can show good discrimination and calibration in a certain extent.the combined model can improve the predictive efficiency and can be used for preoperative prediction and diagnosis of bone metastasis of prostate cancer,which may play a reference role in the choice of treatment and prognosis of patients with PCa. |
PDF Full Text Request