Automated 3D Brain Tumor Segmentation using Attention U-Net with Multi-Task Regularization on Multiparametric MRI

Abstract

In order to make accurate distinctions within gliomas - specifically necrosis, edema, and enhancement - accurate segmentation of these regions from multiple MRI types is required. This task has proven difficult because of how heavily class imbalance affects voxels in this imaging modality, combined with its high degree of morphologic variability. In response to this challenge, a new compact 3D attention-based U-Net is developed that uses multi-task regularization, which is used for the volumetric segmentation of gliomas using data from the BraTS 2023 competition. By applying attention gates at each skip connection in the proposed model, non-tumor-related activations are suppressed before the output of the decoder is concatenated together. Additionally, the addition of a lightweight binary auxiliary classification layer after the first bottleneck in the model can act as an inductive bias to prevent the encoder from collapsing into shortcut texture features based solely on local information when faced with severe class imbalance. The model was tested on a held-out set of 251 patients, resulting in Dice similarity coefficients of 0.927 (whole tumor), 0.877 (tumor core), and 0.831 (enhancing tumor). Furthermore, the median HD95 found for whole tumor boundary localization was 1.57 mm. The model has approximately 5.64 million trainable parameters and performs inference in approximately 0.18 s per patient on a single GPU. This represents a 70% decrease in the number of trainable parameters compared to standard 3D U-Net architectures performing similarly well on segmentation tasks, thus making it suitable for use in resource-constrained deployment settings.