HSTViT: A Hybrid Approach of Vision Transformers and Superimposing For Brain Stroke Classification and Lesion Area Quantification

Accurate diagnosis of brain stroke-like Intracerebral Hemorrhagic Stroke (ICH-S) classification and area quantification of the stroke are tremendously significant for physicians to focus on precise points of the brain and apply the accurate treatment to patients. This paper proposes a hybrid approach called HSTViT that combines a modified vision transformer (MViT) model and a superimposing model for brain stroke classification and lesion area quantification. The MViT model is used for stroke image classification, and the superimposing model is utilized for stroke lesion detection and area quantification. The stroke dataset is collected from the publicly available dataset Kaggle. The dataset contains non-stroke (NS) and ICH-S classes. There are a total of 1850 raw image samples. Thereafter, image preprocessing and augmentation techniques are utilized to enhance the dataset. Later, the dataset is split into three subsets, in which 80% of the samples are exploited for training, 20% are exploited for validation, and 100 samples are considered for testing the model. Furthermore, the proposed HSTViT model is compared with three conventional machine learning models and two deep convolutional neural network (DCNN) models for ensuring the model's efficacy. The HSTViT model performed better than other models and obtained an overall accuracy of 98.78%. The class-wise performance of the model is evaluated and shows 100% accuracy for the NS class and 98.43% accuracy and 98.33% recall for the ICH-S class. Besides, the stroke lesion is detected by the model and quantifies the stroke lesion area. However, the performance of the HSTViT model ensures that the model is suitable for biomedical applications that assist the physicians in detecting the quantification of ICH stroke lesions.