Blood Clot Classification in Ischemic Stroke Patients using ESAFH-Net with Explainable AI.

Accurate classification of blood clots in ischemic stroke patients is vital for effective treatment planning. Nonetheless, difficulties like subtle visual differences and complicated clot structures make manual diagnosis time-consuming and error-prone. This article introduces an innovative model named the Efficient Shuffle Attention Forward Harmonic Network (ESAFH-Net) for classifying blood clots in ischemic stroke patients. Unlike existing blood clot classification approaches, the proposed ESAFH-Net uniquely integrates shuffle attention with harmonic analysis within an EfficientNet backbone to capture both local and global clot characteristics. Furthermore, this study is among the first to incorporate Explainable Artificial Intelligence (XAI) for interpretable classification of Cardioembolic (CE) and Large Artery Atherosclerosis (LAA) clots using whole-slide digital pathology images. Firstly, the input whole slide digital pathology image is obtained from the given dataset. The acquired image is subjected to enhancement using Contrast Limited Adaptive Histogram Equalization (CLAHE) to enhance visual quality. Next, blood clot segmentation is performed using a Recurrent Residual Convolutional Neural Network based on U-Net (R2U-Net), with a log loss function. Thereafter, morphological features and Gray-Level Co-Occurrence Matrix (GLCM) features are mined. Finally, blood clot classification is performed using the proposed ESAFH-Net, enabling differentiation between CE and LAA types. Moreover, ESAFH-Net is a hybrid model that incorporates EfficientNet, Shuffle Attention Network (SA-NET), and harmonic analysis. Furthermore, the interpretability of the proposed technique is validated using XAI. The efficacy of ESAFH-Net is validated employing evaluation measures, achieving a True Positive Rate (TPR) of 92.885%, an accuracy of 92.595%, and a False Positive Rate (FPR) of 7.617%.