FPUNet: A Multi-Level Residual Fractional Domain Transformer Network for Ischemic Stroke Image Segmentation

Abstract

Due to the fact that ischemic stroke patients comprise 60%–70% of all stroke cases, coupled with the long examination time and narrow treatment window, along with the high requirement for clinicians' experience, an accurate and rapid ischemic stroke lesion segmentation algorithm can provide clinicians with valuable assistance in the diagnosis and treatment of stroke patients, which is of great clinical significance. This paper proposes a Fractional Perspective U-Net (FPUNet), which offers a novel perspective for observing lesion features between the spatial and frequency domains, allowing for a more prominent examination of these features. Traditional spatial or frequency domain analysis restricts the observation of signals to two separate angles, making it difficult to simultaneously analyze from both perspectives; this can lead to the oversight of important signal characteristics. In contrast, the fractional domain offers a balance between time and frequency, facilitating the analysis of signals across different scales. This multi-scale perspective enables the capture of details that may be overlooked in pure time or frequency domains. It allows for a more effective extraction of details and texture information from medical images, thereby accurately delineating the edges of stroke regions and providing clearer boundaries for pathological areas, improving the separation of lesions from the background. FPUNet is designed with a multi-level residual structure incorporating a multi-head attention mechanism based on the fractional domain, alongside a variant of convolutional neural network whose layers are tailored to the number of feature map channels for effective channel feature extraction. This innovative approach aims to address the challenges posed by the intricate nature of stroke, ultimately assisting clinicians in the diagnosis and treatment of stroke patients. The proposed method demonstrates superior performance over state-of-the-art models in both accuracy and segmentation efficacy, achieving Dice coefficients of 64.36%, 63.02%, and 86.11% on the AISD, ATLASv2.0, and ISLES22 datasets, respectively.