Image Classification of Ischemic Stroke Blood Clot Origin using Stacked EfficientNet-B0, VGG19 and ResNet-152

Stroke continues to be the second-leading cause of mortality globally. Over 700,000 Americans suffer from an ischemic stroke every year as a result of a blood clot clogging a brain artery. The chances of the patient surviving a second stroke where recurrent strokes account for 23% of all incidents, are decreased. The motivation of this work is to help physicians better understand the causes of blood clots in fatal strokes, making it simpler for doctors to recommend the most effective post-stroke therapeutic management and lowering the risk of a second stroke. Whole slide digital pathology images capture the blood clot of the patient who has suffered from ischemic stroke. The objective of this work is to classify the blood clot origin in the case of ischemic stroke. Its a binary classification problem where the classes are Cardioembolism and Large artery atherosclerosis. Deep learning models can be applied to classify the stroke origin efficiently. This work proposes a stacked deep learning model with VGG19, ResNet-152, EfficientNet-B0 to classify the stroke origin. Experiments have been conducted on Mayo Clinic - Strip AI dataset from Kaggle. Comparisons have been made with these individual models and also combinations of these models and found that the proposed model is achieving lower loss value.