Stroke disease classification from computed tomography images using Inception Harmonic LeNet and wavelet-...

Stroke is a leading cause of mortality, making prompt and precise diagnosis essential for effective treatment. Computed Tomography (CT) screening is crucial in identifying stroke types, particularly ischemic and hemorrhagic strokes. Existing automated methods lack the accuracy and consistency required for reliable stroke diagnosis. Therefore, a novel Inception Harmonic LeNet (InHLeNet) approach is devised for stroke disease classification. Initially, CT scans are collected and subjected to preprocessing, which is done using guided filtering and an improved Non-Subsampled Shearlet Transform (NSST) threshold. The filtered images are then segmented using the Dimension fusion U-Net (D-UNet). Subsequently, augmentation is performed by local augmentation and self-augmentation, where local augmentation introduces localized variations within each CT image, and self-augmentation generates feature-guided transformations of lesion regions. Further, the wavelet transforms with Symmetrically Weighted Local Gradient Pattern (Wavelet-SWLGP) features are extracted. Lastly, stroke disease is classified using InHLeNet, which merges InceptionV3Net, LeNet, and Harmonic analysis. The performance of InHLeNet is assessed using several evaluation metrics, including accuracy, True Positive Rate (TPR), True Negative Rate (TNR), and Matthews Correlation Coefficient (MCC). The results attained using the InHLeNet model is accuracy of 96.888%, TNR of 96.381%, MCC of 96.777%, and TPR of 97.988%, with image size, highlighting its effectiveness.