Developing an Advanced Deep Learning-based MR Image Framework for Brain Stroke Segmentation and Classification with Novel Activation Function

Stroke is considered as one of the most prevalent causes of death and disability for the humans, although it is preventable and treatable. Earlier stroke detection and treatment management helps in enhancing the clinical outcomes, thereby significantly minimizing the risk of disease. Because machine learning techniques are recently implemented to detect strokes and it gains a lot of attention. However, manual feature engineering and limited generalization are common problems with conventional machine learning models, which might lower diagnostic reliability. In order to address limitations in conventional models, this work presents a sophisticated deep learning-based stroke prediction framework using the Magnetic Resonance Imaging (MRI) and provides specialized and flexible diagnostic guidance. The developed stroke detection system begins by collecting the required MR images in the benchmark sources. Further, the gathered MR images are fed to the stroke lesion segmentation using the developed Region Masked Attention-based Multi-Dilated Inception Unet++ (RMA-MIUnet++), which is accurately focus on stroke-affected regions. The segmented images are acquired as the outcomes from the proposed RMA-MIUnet++ model. These segmented images are further classified in the developed Efficient InceptionV3 with Novel Activation Function (EIV3-NAF)-based stroke classification model. The results achieved by the developed EIV3-NAF model clearly show enriched performance, whereas the model's performance was tested and compared with baseline methods to prove how well the model works in classifying the strokes.