Hyperparameter-optimized Cross Patch Attention (CPAM) UNET for accurate ischemia and hemorrhage segmentation in CT images

Abstract

Stroke represents a critical medical condition with the potential for substantial brain damage and severe complications. Timely identification of stroke is imperative to minimize harm and enhance patient prognosis. Computed tomography (CT) scans serve as a standard tool for identifying stroke presence and location due to their capacity to offer intricate brain images. Nonetheless, accurately classifying stroke types and delineating their boundaries remain challenging due to dataset limitations and algorithmic constraints. This study focuses on the precise classification of stroke types from contrast-agent-free brain CT images. A meticulously curated dataset of stroke images was collaboratively assembled with domain experts. Hyperparameter optimization techniques were applied to evaluate deep learning classification models. The resultant stroke segmentations were visualized via an enhanced U-Net model, which integrates the Cross Patch Attention Module (CPAM) to elevate segmentation accuracy. Our investigation meticulously examines a spectrum of classification models with the principal goal of distinctly discerning diverse stroke types. Benchmarking the renowned VGG16 model for image classification yielded a commendable 94% accuracy in identifying stroke types. However, the CPAM-Unet model exhibited superior performance, achieving an impressive 95% accuracy. Notably, the CPAM-Unet model demonstrated an Intersection over Union (IOU) of 88%, a metric conventionally associated with segmentation tasks. This performance hallmark underscores its adeptness in discriminating between ischemic and hemorrhagic strokes. This study showcases the potential of deep learning in stroke detection using contrast-agent-free CT images. The outcomes underscore the effectiveness of the CPAM-Unet model in accurately classifying strokes, surpassing other classification models. By harnessing deep learning capabilities and incorporating the CPAM module, heightened segmentation accuracy for ischemic and hemorrhagic strokes can be achieved. These findings contribute to the domain of stroke diagnosis, underscoring the need for further research to advance early detection and enhance patient outcomes. The results presented in this study hold promise for refining clinical practices and optimizing stroke patient management, thereby warranting attention from both the scientific community and medical practitioners.