A hybrid framework combining enhanced fields and consistent spatial features for stroke lesion segmentation in MR images

Abstract

Stroke lesion segmentation plays a critical role in clinical diagnosis, prognosis and rehabilitation treatment. However, this task remains highly challenging due to inter-subject differences, shape variations, noise and artifacts in MR images, which hinder the accurate extraction of stroke regions. This work proposes a novel hybrid framework for accurate stroke lesion segmentation from MR images termed HFFS-net. To enhance spatial consistency between the segmented stroke regions and the original ones in MR images, we introduce an adversarial architecture as an auxiliary encoder, which extracts high-dimensional distribution features of stroke lesions, ensuring improved consistency between the spatial distributions of original MR images and segmented outputs. To further capture fine-grained details of stroke lesions, we enhance the receptive fields of kernels in convolution framework, enabling robust extraction of representative features from target stroke lesions. Additionally, we design a multi-dimensional framework with hierarchical enhanced fields to comprehensively aggregate diverse stroke-specific information from MR images. The proposed model was rigorously evaluated on datasets comprising stroke lesions of varying sizes to comprehensively assess its segmentation performance and generalization capability. Experimental results demonstrate that our framework outperforms existing segmentation methods, attaining a Dice score of 84.33% and a Recall of 85.44%. These results confirm that the proposed framework is capable of effectively capturing diverse and discriminative features, thereby enabling accurate stroke lesion segmentation.