H Li^1*, T Martin-Carreras^2* , K Cooper³ , J Yang⁴ , R Sebro^5# , Y Fan^6# , (1) University of Pennsylvania, Philadelphia, PA, (2) University of Pennsylvania, Philadelphia, PA, (3) University of Pennsylvania, Philadelphia, PA, (4) UT MD Anderson Cancer Center, Houston, TX, (5) University of Pennsylvania, Philadelphia, PA, (6) University of Pennsylvania, Philadelphia, Pennsylvania (^*Li and Martin-Carreras share the first authorship, ^#Sebro and Fan share the last authorship)

• 
  H Li

(Wednesday, 8/1/2018) 10:30 AM - 11:00 AM

Room: Exhibit Hall | Forum 5

Purpose: Myxoid tumors are tumors that pose great diagnostic challenges for radiologists. Myxomas are benign pauci-cellular, bland myxoid tumors with no propensity to metastasize, but may locally recur. Myxofibrosarcomas are malignant myxoid tumors that are extremely heterogeneous histologically and on MRI, and have the propensity to metastasize and are on the other end of the spectrum. In this study, we evaluated the performance of radiomic features extracted from magnetic resonance images (MRI) for distinguishing myxomas from myxofibrosarcomas.

Methods: This study was performed based on a T1 MRI dataset of 56 patients, consisting of 29 patients with myxoma and 27 patients with myxofibrosaroma tumors. From T1 scan of each patient, we extracted 89 radiomic features including shape features, intensity statistics, Gray level co-occurrence matrix, Gray level run-length matrix, and Gray level size zone matrix based features. Random forests based classifiers were built upon the radiomic features for distinguishing myxoma from myxofibrosarcoma tumors. Classification performance of the classifiers built upon radiomic features was compared with those built upon shape features only. The number of trees and minimum leaf size of the random forests classifiers were set to 200 and 3 respectively.

Results: The classifiers were validated using a leave-one-out cross-validation, and classification accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) were used to evaluate the classification performance. The accuracy, sensitivity, specificity, and AUC obtained by the classification model built upon radiomic features were 0.839, 0.815, 0.862, and 0.880 respectively. Our classification model outperformed the classification model built upon the shape features, with the performance measures as 0.804, 0.778, 0.828, and 0.858 for accuracy, sensitivity, specificity, and AUC respectively.

Conclusion: Experimental results on T1 MRI data have demonstrated that radiomic features could provide more discriminative information for distinguishing myxoma from myxofibrosarcoma tumors compared to conventional shape based measures used.

Keywords

CAD, MRI

Taxonomy

IM- MRI : Quantitative Imaging/Analysis

Contact Email