Room: ePoster Forums
Purpose: To improve accuracy of brain tumor classification using MRI radiomics data.
Methods: 62 MRI scans were collected under IRB exemption to include 11 embryonal tumor patients (medulloblastoma) and 51 patients with non-embryonal tumors. Tumors were segmented at 4 levels in each of T1 + Gadolinium, T2, T2 FLAIR, and Apparent Diffusion Coefficient (ADC) series. All series were acquired in the axial orientation and the 4 tumor levels matched between the series. Radiomics features were calculated using 3D Slicer (www.slicer.org). Three data sets were generated from the 16 feature sets for each patient. The first data set averaged the radiomics features from the four levels in each sequence and concatenated into one feature vector per patient. In the second set, the features at each level were concatenated giving 4 feature vectors per patient. In the third data set, all permutations of feature sets were created by concatenating features from different levels across the 4 series, generating 256 feature vectors per patient. Each of the 256 vectors contains (in the same order) T1, T2, FLAIR, and ADC features. A neural network consisting of two 8-node layers separated by a dropout layer was created in Tensorflow. The network was trained and evaluated 5 times for each data set using leave-one-out cross validation. Performance metrics were calculated for each case.
Results: For the averaged feature data, the accuracy, sensitivity, specificity, negative predictive value, and positive predictive values were 71.6%, 0.15, 0.84, 0.16, and 0.82, respectively. For the second data set, the values were 75.6%, 0.28, 0.86, 0.30, and 0.85, respectively. For the fully augmented data, the results were 75%, 0.31, 0.85, 0.31, and 0.86, respectively.
Conclusion: Data augmentation is helpful in preventing overlearning in machine learning application using neural networks. The technique described here modestly improved network performance in this limited-data setting.