A Lightweight Entropy–Curvature-Based Attention Mechanism for Meningioma Segmentation in MRI Images

Meningiomas are a common type of brain tumor. Due to their location within the cranial cavity, they can potentially cause irreversible damage to adjacent brain tissues. Clinical practice typically involves surgical resection for tumors that provoke symptoms and exhibit continued growth. Given the variability in the size and location of meningiomas, achieving rapid and precise localization is critical in clinical practice. Typically, meningiomas are imaged using magnetic resonance imaging (MRI), which produces 3D images that require significant memory resources for the segmentation task. In this paper, a lightweight 3D attention mechanism based on entropy–curvature (ECA) is proposed, which significantly enhances both parameter efficiency and inference accuracy. This attention mechanism uses a pooling method and two spatial attention modules to effectively reduce computational complexity while capturing spatial feature information. In terms of pooling, a tri-axis pooling method is developed to maximize information retention during the dimensionality reduction process of meningioma data, allowing the application of two-dimensional attention techniques to 3D medical images. Subsequently, this mechanism utilizes information entropy and curvature filters to filter noise and enhance feature information. Moreover, to validate the proposed method, the meningioma dataset from West China Hospital’s Department of Neurosurgery and the BraTS2021 dataset are used in our experiments. The results demonstrated superior performance compared to the state-of-the-art methods.