Hybrid Radiomic–Deep Feature Fusion for Pediatric Brain Tumor MRI Analysis

Abstract

Pediatric brain tumors are severe and may be life-threatening diseases in children and the precise analysis of magnetic resonance imaging (MRI) is the key to proper diagnosis, treatment planning and assessing prognosis. Nevertheless, these tumors are highly variable in terms of shape, size and patterns of intensity in different patients so that automated segmentation and classification is difficult. This challenge is further complicated by the fact that pediatric MRI data are not readily available which can decrease the resilience and generalization capacity of deep learning models. Even though convolutional neural networks (CNNs) have demonstrated high performance in the analysis of brain tumors, their main justification is based on the automatically obtained representations that might not be interpretable and stable under a small-data investigation. To overcome these shortcomings, this paper has developed a hybrid radiomic-deep feature fusion model to examine pediatric brain tumor MRI data based on the BraTS-PED 2024 dataset. The proposed method uses 3D U-Net model to perform multi-class tumor segmentation, and then handcrafted radiomic features are extracted on the segmented tumor regions, and deep features are ex- tracted on a CNN classifier. The complementary feature representations are combined to enhance the performance of classification. The experiment findings show that edema (ED), non-enhancing tumor (NET), and enhancing tumor (ET) areas have Dice similarity coefficients of 0.8509, 0.7182 and 0.7058, respectively, with the background segmenta- tion having a coefficient of 0.85. To classify tumors, the deep learning-based model gave a 97.00 percent valida- tion accuracy, whereas a radiomics-based Random Forest model gave a 98.0 percent accuracy. The results have shown, however, that the combination of radiomic and deep features offers free information, increasing the strength, explanatory power, and general validity in the segmentation and categorization of pediatric brain tumors MRI.

Keywords: Pediatric brain tumor, MRI, Radiomics, Deep learning, Feature fusion, 3D U-Net, Tumor segmenta- tion, Tumor classification, Hybrid model, Machine learning