Bias Corrected Twin Squeeze-and-Excitation Attention Enhanced UNet for brain tumor segmentation

Brain tumor segmentation is crucial in the context of deep learning-based medical image analysis, where accurate delineation aids in detecting abnormalities, treatment planning, and monitoring therapeutic outcomes for brain cancer. Despite the success of deep learning algorithms in medical image segmentation, challenges remain in capturing long-range dependencies, extracting relevant features, and addressing intensity variations across different imaging modalities. In this paper, we propose a novel deep learning architecture, Bias-Corrected Twin Squeeze-and-Excitation Attention Enhanced UNet (BC-TSEA-UNet), which integrates twin squeeze-and-excitation (SE) attention blocks into the UNet backbone to tackle these challenges. Unlike standard SE blocks, the twin SE configuration applies dual channel recalibration at multiple semantic levels, thereby capturing both shallow and deep contextual dependencies more effectively. The motivation behind incorporating SE blocks is to enhance feature recalibration, allowing the model to better focus on both local and global contextual information critical for accurate segmentation. Furthermore, a bias correction mechanism is employed during preprocessing to mitigate intensity non-uniformities in MRI scans, ensuring more consistent data representation across modalities. The proposed architecture is extensively evaluated on the BraTS 2019, BraTS 2020, and BraTS 2023 datasets, demonstrating average absolute improvements of 0.1112, 0.1339, and 0.1986 in Dice scores and 0.1601, 0.1396, and 0.1847 in mean intersection-over-union (IoU) on tumor subregions compared to the baseline UNet model, respectively. By emphasizing salient features and mitigating bias, BC-TSEA-UNet significantly improves feature representation, leading to more accurate and reliable tumor delineation across multiple datasets. • Twin SE module is embedded into UNet backbone for precise Brain Tumor segmentation. • Bias-Field correction is imbibed in preprocessing step to negate intensity variation. • Model is evaluated in two phases on MICCAI BraTS 2019, 2020 and 2023 datasets. • Comparable Dice and mean IoU scores achieved, attributed to twin SE and UNet fusion.