A Hybrid Model for Ischemic Stroke Brain Segmentation from MRI Images using CBAM and ResNet50-UNet

Ischemic stroke is the most prevalent type of stroke and a leading cause of mortality and long-term impairment globally. Timely identification, precise localization, and early detection of ischemic stroke lesions brain are critical in healthcare. Various modalities are employed for detection, and magnetic resonance imaging stands out as the most effective. Different magnetic resonance imaging techniques have been proposed for the detection of ischemic stroke lesion tumors, allowing for image uploading and visualization. Automated segmentation of ischemic stroke lesions from magnetic resonance imaging images has an important role in the analysis, prognostic, diagnosis, and clinical treatment planning of some neurological diseases. Recently, computer-aided diagnosis systems based on deep learning techniques have demonstrated significant promise in medical image analysis, particularly in multi-modality medical image segmentation. Automated segmentation is a difficult task due to the enormous quantity of data provided by magnetic resonance imaging and the variation in the location and size of the lesion. In this study, we develop an automated computer-aided diagnosis system for the automatic segmentation of ischemic stroke lesions from magnetic resonance imaging images using a Convolution Block Attention Module (CBAM) and hybrid UNet-ResNet50 model. The UNet model is integrated into the architecture, and the ResNet50 backbone is pre-trained to enhance feature extraction. CBAM block is a model applied in this approach to extract the most effective feature maps. The proposed approach is evaluated on the public Ischemic Stroke Lesion Segmentation Challenge 2015 dataset, arranged into weighted-T1(T1), weighted-T2(T2), FLAIR, and DWI sequences. Experimental results demonstrate the efficacy of our approach, achieving an impressive accuracy value of 99.56%, a precision value of 97.12%, and a DC of 79.6%. Notably, our approach outperforms other state-of-the-art methods, particularly in terms of accuracy values, highlighting its potential as a robust tool for automated ischemic stroke lesion segmentation in magnetic resonance imaging.