EnigmaNet: A Novel Attention-Enhanced Segmentation Framework for Ischemic Stroke Lesion Detection in Brain MRI

Abstract

Segmentation of lesions is crucial for the detection and treatment of ischemic stroke. The aim of this work is to develop a robust and highly accurate framework for lesion stroke detection in brain Magnetic Resonance Imaging (MRI). The paper propose a novel deep learning model, named EnigmaNet, for the segmentation of ischemic stroke lesions in Fluid-Attenuated Inversion Recovery (FLAIR) and Diffusion Weighted Imaging (DWI) images. EnigmaNet use novel Genesis-k blocks and dual attention mechanism in the encoder and decoder blocks of the architecture. A modified Weighted Focal-Tversky-Dice (wFTD) Loss is used for improved performance. The model is validated on the ISLES-2015 public dataset. EnigmaNet showed a Dice score of 0.8965, sensitivity of 0.8776 and specificity of 0.9866 for the FLAIR test images. Dice score of 0.8423, sensitivity of 0.8452 and specificity of 0.9754 were obtained for DWI images. The segmentation results of EnigmaNet shows an improvement in Dice score of about 32% over U-Net-sharp, 41% over FCN-8 and 10% over Attention U-Net. A region-based comparison of segmentation results of EnigmaNet highlighted its ability to detect fine lesions accurately in multiple vascular territories in brain, thereby signifying its robustness to segment lesions of diverse size, shape and location. The proposed model showed improved results as compared to the state-of-the-art techniques. EnigmaNet model is thus a promising approach for accurate and robust segmentation of ischemic stroke lesions.

Keywords

• Magnetic resonance imaging
• deep learning
• image segmentation
• ischemic stroke lesion
• FLAIR
• DWI