Abstract
58
Objectives: Osteosarcoma (OS) is extremely heterogeneous, and it is manifested in in children and adolescents between the ages of ten and twenty-five years. Radiomics model can be clinically better feasible prediction model for treatment decision using relation to a given tumor outcomes as prognostic factors. In this study logistic regression based on machine learning algorithm was developed and evaluated using metabolic imaging phenotypes.
Methods: A database of eighty-two consecutive patients was retrospectively retrieved. Of the patients, 26.83% had metastasis (lung metastasis: 72.7%, bone metastasis: 27.3%) after completion of neoadjuvant chemotherapy and surgery in five years or seven years (one patient). Their pre-treatment 18F-FDG PET/CT scans were used for radiomics model. The tumor region was drawn using a semi-automated segmentation method with threshold SUV of 2.0 in three-dimensional (3D) images. Forty-five features were extracted from the segmented tumors. The incorporation of features into multivariable models was selected by Spearman correlation, stepwise elimination method based on Akaike's An Information Criterion (AIC), ANOVA test, and odds ratio. The logistic regression model based on 10-fold cross-validation method was set up by the selected features, validated by training set (60% of total patients), and tested by independent test set (40% of total patients). The performance was evaluated in terms of area under curve (AUC), sensitivity, specificity, accuracy, and precision computed based on receiver-operating-characteristic (ROC) curves and confusion matrix.
Results: Two features, SUVmax (SUV maximum) and GLZLM (Gray-Level Zone Length based on intensity-size-zone Matrix) -SZLGE (Short-Zone Low Grey-Level Emphasis) were chosen by the selection criteria: Spearman correlation (p-value ≤ 0.05), Friedman’s ANOVA test (p-value < 0.05), odds-ratio (> 1.0), and multicollinearity (VIF < 4.0). These features are used for the multivariable logistic regression model. The trained and validated multivariable logistic model based on probability of endpoint (P) = 1/ (1+exp (-Z)) was Z= -1.23 + 1.53[asterisk]SUVmax + 1.68[asterisk]GLZLM_SZLGE with significant p-values (SUVmax: 0.0462 and GLZLM_SZLGE: 0.0154). The final multivariable logistic model achieved an area under the curve (AUC) receiver operating characteristics (ROC) curve of 0.80 which showed better result than those of individual group of features (SUVmax: 0.38, GLZLM-SZLGE: 0.56 ). The accuracy was 0.81, and the specificity was 0.88. The sensitivity (0.63) was lower than other values, even though it was fair. This occurred because of the insufficient number of metastasis cases -only eight.
Conclusions: The selected two features, SUVmax and GLZLM_SZLGE from metabolic imaging phenotypes are independent predictors of metastasis risk estimation. The multivariable model developed using them could improve patient outcomes by allowing aggressive treatment in patients identified with high metastasis risk probability.