The New Double-Branch 3D U_i - Net: Design, Development, and Validation of Brain Tumor Clinical Features

Brain tumors represent a heterogeneous and clinically challenging class of pathologies affecting the central nervous system, where early and accurate characterization is essential for patient management. Magnetic Resonance Imaging (MRI) is the current gold standard for non-invasive tumor assessment, providing multi-contrast information through sequences such as T1, T1-CE, T2, and FLAIR. To contextualize this study, a systematic literature review was conducted to summarize recent state-of-the-art studies that employ Deep Learning (DL) methods for MRI-based feature extraction and tumor segmentation. The aim of this work is to design, develop, and evaluate the novel Double-Branch 3D U_i - Net for directly regressing clinically relevant quantitative tumor features from preprocessed brain MRI images. The network follows a symmetric encoder-decoder structure inspired by U-Net but integrates an additional regression branch directly connected to the bottleneck, enabling the simultaneous extraction of 14 radiomic-inspired descriptors, including Center of Mass (CoM), bounding-box coordinates, spatial extents, and directional kurtosis, directly from multimodal MRI without relying on handcrafted radiomic pipelines. The model was trained on the BraTS 2020 dataset and tested on BraTS 2017 to assess generalization, with regression performance evaluated using the relative error for each feature. Quantitative results show that the network achieves stable and accurate predictions of global geometric descriptors, with relative errors ranging from 16% to 30% for CoM coordinates and bounding box, while higher-order boundary-dependent features, such as kurtosis and minimum and maximum distances from the CoM present larger variability, with average relative errors of 30% - 34%, reflecting the intrinsic irregularity of glioma margins. These findings confirm the model’s ability to extract clinically meaningful morphological information directly from the latent representation. Overall, the proposed Double-Branch 3D U_i - Net provides an automated and reproducible framework for MRI-based brain tumor characterization, combining precise tumor localization with quantitative feature extraction in a single end-to-end pipeline. This integrated strategy reduces dependence on manual annotation and handcrafted radiomics pipelines, enhances reproducibility and diagnostic accuracy, and provides a foundation for future applications in treatment planning, longitudinal monitoring, and personalized neuro-oncology interventions.