Intracranial Hemorrhage Classification From CT Scan Using Deep Learning and Bayesian Optimization
Shifat E. Arman, Sayed Saminur Rahman, Niloy Irtisam, Shamim Ahmed Deowan, Md. Ariful Islam, Saadman Sakib, Mehedi Hasan
Abstract
Intracranial hemorrhage is a medical condition that involves bleeding within the skull or brain tissue. It has mainly five subtypes - epidural hemorrhage, subdural hemorrhage, subarachnoid hemorrhage, intraparenchymal hemorrhage, and intraventricular hemorrhage. In order to ensure a successful outcome for a patient, a timely and accurate detection of intracranial hemorrhage is crucial. Despite this, there is a shortage of radiologists, especially in rural areas, which can lead to a delay in diagnosis. In this work, we proposed an automatic way of diagnosing intracranial hemorrhage from a CT scan. We have optimized the DenseNet architecture using Bayesian Optimization (BO) to diagnose intracranial hemorrhage effectively. Using BO, we identified the optimal learning rate, optimizer, and number of dense nodes for the DenseNet architecture. Our proposed model can analyse whether hemorrhage is present in a CT scan and, if it is present, what its subtype is. The optimized DenseNet model was able to achieve a very high accuracy of 98.02% on the test set. By ensuring accurate and reliable diagnoses, our method will assist doctors in making better-informed decisions and providing better care for their patients.