Enhancing Stroke Detection in Brain CT Images with Residual Attention Networks

Stroke is becoming one of the most common diseases in modern society. Stroke caught in the early phase has a higher potential for recovery as it can be treated before the worsening of the patient's condition. Detection through brain CT images with deep learning has a favorable impact. Residual Attention Network (RAN) architecture is proposed to perform stroke detection. RAN builds upon the regular ResNet architecture by incorporating an attention layer within each residual block to extract meaningful information. RAN combines the advantages of residual connections and attention mechanisms. RAN is able to prevent vanishing gradient problem while focusing on the most important features. RAN yields an accuracy of 96.12% surpassing the regular ResNet with an accuracy of 94.40%. RAN shows great potential, especially in fields that require high accuracy and efficient feature extraction. This paper explores the effectiveness of Residual Attention Networks and their potential applications in areas requiring robust and accurate feature extraction.