NIMG-61. USE OF TEXTURAL RADIOMIC MAPS IN A 3D CONVOLUTIONAL NEURAL NETWORK FRAMEWORK CAN AUGMENT GLIOMA LESION SEGMENTATION

Abstract

OBJECTIVE: Gliomas, one of the most common primary brain malignancies, exhibit phenotypically heterogeneous histological sub-regions (edema ED, necrotic core N, enhancing E, non-enhancing NE), each containing relevant diagnostic and prognostic information. Segmentation of these subregions in multimodal MRI is a major challenge, and usually performed manually. In this work, we hypothesize that radiomic features capturing sub-visual heterogeneity information on routine MRI (T1c, T2w, FLAIR), in conjunction with a Convolutional Neural Network (CNN) classifier will improve identification of lesion boundaries and hence improve segmentation, as compared to using intensity alone. METHODS: 221 high grade glioma studies were acquired from the BraTS 2016 challenge. After sequence co-registration, all cases were skull-stripped and normalized to correct for intensity inhomogeneity. Ground truth sub-compartment delineations were obtained from clinical experts. Four gray-level co-occurrence features were extracted from each MRI protocol. A 3D CNN was trained on N=154, validated on N=33, and tested on N=34 using original intensities and extracted feature maps. The architecture comprised eleven layers: eight convolutionalpooling layers followed by two fully connected layers and one classification layer. Both fully connected layers had 150 neurons connected to the two final neurons to determine voxel sub-type. We compared the performance of our model against a standard intensity-trained model and other similar intensity-trained pipelines from the challenge. Dice Similarity Coefficient (DSC) scores were used to evaluate segmentation performance. RESULTS: Using radiomic-CNN, there was an improvement in the DSC scores (0.82 and 0.80) for non-enhancing and enhancing tumor respectively as compared to the intensity-only CNN (0.79 and 0.79). In comparison to 18 segmentation methods, our method performed better (0.90 vs. 0.84-0.87) in segmenting tumor core (N+E+NE) and enhancing region (0.80 vs. 0.72-0.76), and comparable in segmenting whole tumor (ED+N+E+NE). CONCLUSION: Our results suggest that radiomic features in conjunction with 3D CNNs can augment lesion segmentation performance.