Hybrid Attention and Learnable Thresholding for 3D Brain Tumor Segmentation

Abstract

Rapid and accurate segmentation of tumors in medical images is of primary importance for cancer diagnosis and treatment designs. Though advancements have been made with autoencoding 3D images and U-Net-like architectures, the segmentation boundary precision problem is still persistent with complex, oddly shaped tumors under constrained input conditions. Advancing boundary precision remains one of the field’s most important open challenges. This study proposes an improvement to the 3D autoencoder by incorporating a hybrid model. We improve feature extraction through the addition of a hybrid attention (HA) mechanism, which combines channel-wise and learned query structural attention. We also include a differentiable learnable thresholding (DLT) layer, which provides data-driven segmentation boundary control and segmentation boundary refinement for segmentation. The model was evaluated on the Brain Tumor Segmentation (BraTS) 2020 dataset and achieved a 92.0 dice score on the whole tumor. The defined metrics serve as robust evidence for accuracy and demonstrate the improvement of the proposed model over existing state-of-the-art benchmarks. Increased precision and accuracy offered by the proposed system surpasses the current computerassisted diagnostic systems for medical imaging. This study pushes the boundary of existing knowledge as it helps to resolve the challenging issue of pinpointing the location of tumors in 3D medical volumes.