MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Musculoskeletal Tumor Classification On T2-Weighted MRI Using Probability Fusion Convolutional Neural Network and Support Vector Machine

L Chen*, S Fisher , A Rodriguez , M Folkert , A Chhabra , S Jiang , J Wang , UT Southwestern Medical Center, Dallas, TX

Presentations

(Wednesday, 7/17/2019) 7:30 AM - 9:30 AM

Room: 225BCD

Purpose: Classification of musculoskeletal lesions from T2-weighted MR imaging alone proves to be a difficult task for radiologists due to the wide variability of imaging features used for diagnosis. Machine learning may allow for mining quantitative imaging features for potentially increased accuracy of diagnosis. Radiomics and deep learning methods are two main promising strategies using imaging features to predict lesion malignancy. The purpose of this study is to develop an accurate, reliable and automatic hybrid prediction model by taking advantage of both radiomics and deep learning methods for musculoskeletal tumor classification.

Methods: T2-weighted MR images from a cohort of 45 patients (21 benign & 24 malignant) with histologically verified musculoskeletal masses were analyzed and used to train predictive classification models. Additional 43 patients (18 benign & 25 malignant) data were used as an independent validation cohort. We firstly designed a convolutional neural network (CNN) with Alex-net-type architecture using region of interest (ROI) patches including lesion and their surrounding voxels as inputs. The proposed CNN model learns the local and global features automatically by arranging the convolutional and max-pooling layers alternately in the CNN architecture. Additionally, we extracted seven geometrical features, and 819 texture features from the segmented lesions. Relieff feature selection algorithm was performed to prevent information redundancy. Then we used a support vector machine (SVM) to build a predictive model. After obtaining the outputs from the SVM and CNN models, the final output was generated using analytic evidential reasoning (ER) algorithm.

Results: Accuracy (ACC) and area under the curve (AUC) were used to assess the predictive accuracy of the predictive models. The hybrid method achieved an ACC (AUC) value of 0.86 (0.92) while CNN and SVM alone achieved 0.81 (0.88) and 0.84 (0.87), respectively.

Conclusion: The hybrid method generates more accurate prediction results for musculoskeletal tumor malignancy using T2-weighted MRI as compared to SVM or CNN alone.

Funding Support, Disclosures, and Conflict of Interest: US National Institutes of Health (R01 EB020366)

Keywords

Not Applicable / None Entered.

Taxonomy

IM- MRI : Machine learning, computer vision

Contact Email