A Survey on Deep Learning for Intracranial Hemorrhage Detection

Computed Tomography (CT) images of the patient's cranium are widely used in the diagnosis of hemorrhage. They are reviewed by highly skilled professionals to find the existence, location and type of hemorrhage. This process is complex, labor-intensive and requires lot of time. The massive progress in machine learning techniques and availability of high computational power has made it possible to detect such hemorrhages automatically. The paper provides survey of recent deep learning approaches for automatic detection of such acute intracranial hemorrhage and its subtypes. Further the experimental results using windowing technique inspired by practical radiologist's approach and relevant research papers is provided to verify its efficacy as a pre-processing step before feeding DICOM images to the network. Transfer learning approach is used to detect hemorrhage types. The open source Kaggle platform is utilized for computational resources and for carrying out the experimental study. Some key evaluation metrics and loss functions addressing the data imbalance underlying with the problem statement are also presented. The paper concludes with addressing key insights into open issues in the proposed domain.

Automated summary

The study introduces recent deep learning approaches for automatic detection of intracranial hemorrhage, and validates the efficacy of windowing technique as a preprocessing step. Transfer learning is used for hemorrhage type detection, with experimental studies conducted on the Kaggle platform. Key evaluation metrics and loss functions addressing data imbalance issues are also discussed.