Alessandro Gronchi, Luca T. Mainardi, Alfonso Marchianò, Paolo G. Casali, Antonella Messina, Eros Montin, Valentina D.A. Corino
To assess the feasibility of grading soft tissue sarcomas (STSs) using MRI features (radiomics).
Materials and Methods
MRI (echo planar SE, 1.5T) from 19 patients with STSs and a known histological grading, were retrospectively analyzed. The apparent diffusion coefficient (ADC) maps, obtained by diffusion-weighted imaging acquisitions, were analyzed through 65 radiomic features, intensity-based (first order statistics, FOS) and texture (gray level co-occurrence matrix, GLCM; and gray level run length matrix, GLRLM) features. Feature selection (sequential forward floating search) and classification (k-nearest neighbor classifier) were performed to distinguish intermediate- from high-grade STSs. Classification was performed using the three different sub-groups of features separately as well as all the features together. The entire dataset was divided in three subsets: the training, validation and test set, containing, respectively, 60, 30, and 10% of the data.
Intermediate-grade lesions had a higher and less disperse ADC values compared with high-grade ones: most of FOS related to intensity are higher for the intermediate-grade STSs, while FOS related to signal variability were higher in the high grade (e.g., the feature variance is 2.6*105 ± 0.9*105 versus 3.3*105 ± 1.6*105, P = 0.3). The GLCM features related to entropy and dissimilarity were higher in the high-grade. When performing classification, the best accuracy is obtained with a maximum of three features for each subgroup, FOS features being those leading to the best classification (validation set: FOS accuracy 0.90 ± 0.11, area under the curve [AUC] 0.85 ± 0.16; test set: FOS accuracy 0.88 ± 0.25, AUC 0.87 ± 0.34).
Good accuracy and AUC could be obtained using only few Radiomic features, belonging to the FOS class.
Level of Evidence: 4
Technical Efficacy: Stage 2
J. Magn. Reson. Imaging 2017.