DSP-UNet: Dual-Skip Perceiver UNet for Lower-Grade Glioma Segmentation

Accurate segmentation of gliomas in clinical neuroimaging is critical for diagnosis and treatment planning. Yet, integrating Artificial Intelligence (AI) into this domain faces challenges due to the need for large, annotated datasets and the complexities of brain tumor morphology. To address these issues, we propose Dual-Skip Perceiver UNet (DSP-UNet), a novel Deep Learning (DL) model specifically designed for glioma segmentation. DSP-UNet incorporates dual-skip connections with mixed kernel architectures and a shared decoder to enhance feature representation and abstraction. The model efficiently extracts features and performs segmentation with reduced dependence on extensive datasets, overcoming the limitations of traditional Convolutional Neural Networks (CNNs). We evaluated DSP-UNet using glioma images from The Cancer Imaging Archive (TCIA). The model achieved outstanding performance with an accuracy of 99.37%, a Mean IoU of 91.87%, and a precision of 90.85%, outperforming existing state-of-the-art approaches. To enhance model interpretability, we employed saliency maps, which provided insights into the decision-making process. These results demonstrate the potential of DSP-UNet to advance AI-driven glioma segmentation, paving the way for more efficient and accurate neuroimaging solutions with reduced reliance on labor-intensive data annotation.