Purpose: To develop and verify a subregional radiomic-based framework for image segmentation(RFIS).
Methods: RFIS was designed using features extracted from subvolume (svfeatures) created by sliding window(swvolume). Fifty-three svfeatures were extracted from 11 phantom images. Outliers were detected by isolation forest(iForest) and eliminated by specified as the mean value of svfeatures. The percentage coefficient of variation (%COV) was calculated to evaluate the reproducibility of svfeatures. RFIS for the gross target volume (GTV) segmentation from the peritumoral region (GTV with 10-mm margins) was constructed to assess the feasibility of RFIS. A total of 127 lung cancer images (training: validation = 86:41) were enrolled. A test-retest method, correlation matrix, and Mann-Whitney U tests (p<0.05) were used for feature selection. Synthetic minority over-sampling technique (SMOTE) was used to balance the minority group in the training datasets. Support vector machine (SVM) was employed for RFIS construction, which is tuned in the training datasets using ten-fold stratified cross-validation. RFIS performance was evaluated by AUC, accuracy, sensitivity, specificity and the Dice similarity coefficient (DSC).
Results: A total of 30249 phantom and 145008 patient image swvolumes (training: validation =97374:47634) were analyzed. Forty-nine svfeatures had good reproducibility. Forty-five features included five categories that passed test-retest analysis. Thirteen svfeatures were selected for RFIS construction. RFIS had a mean sensitivity of 0.848, a specificity of 0.821, an accuracy of 83.48%, and an AUC of 0.906 with cross-validation. The sensitivity, specificity, accuracy and AUC were equal to 0.762, 0.840, 82.29%, and 0.877 in the external validation dataset. The mean DSC was 0.707±0.093 in training datasets and 0.688±0.072 in external validation datasets.
Conclusion: Reproducible svfeatures can capture quantitative image information difference between swvolumes. RFIS can be applied to swvolume classification, which achieves image segmentation by grouping and merging the swvolume with similar quantitative image information.
CT, Image Analysis, Segmentation