Abstract
This research introduces an advanced UNet architecture for the precise segmentation of stroke lesions in Computed Tomography (CT) brain scans, addressing the critical need for accurate and efficient diagnosis in clinical settings. The literature reveals a spectrum of technologies utilized in detecting brain lesions, ranging from manual segmentation to threshold-based techniques and conventional machine learning methods, each with its limitations in terms of accuracy, time efficiency, and generalization capability. Deep learning techniques, notably the UNet model, have emerged as promising alternatives, offering significant advancements in medical image segmentation due to their ability to learn complex patterns and features from data. Our enhanced UNet model incorporates innovative techniques such as data augmentation, dropout, the Adam optimization algorithm, L2 regularization, and instance normalization to address the inherent challenges in stroke lesion segmentation, including the subtle appearance of lesions and the variability across different cases. The comparison of our approach with traditional methods using the ISLES 2018 dataset demonstrates substantial improvements in key performance metrics such as dice/f1 score, precision, recall/sensitivity, and Jaccard index, highlighting the effectiveness of our model in achieving precise segmentation outcomes. The utilization of these advanced techniques enables our model to outperform existing methods, offering a significant advantage in the rapid and accurate segmentation of stroke lesions. This research underscores the potential of the UNet architecture, augmented with modern deep learning techniques, to set a new benchmark in the field of stroke lesion detection, providing a robust tool for enhancing diagnostic processes and patient care in neurology.