ABSTRACT: Image recognition via convolutional neural networks (CNN) has made giant strides in the last 10 years with the potential of outperforming expert-level human accuracy through adaptive learning and high dimensional feature extraction. Recently, they have gained a central stage in radiological tasks and medical diagnostics.
In this project, we implement learning through a single-stage, end-to-end, convolutional neural network frameworks based on InceptionNet and EfficientNet to address the problem of classifying multiple types intracranial hemorrhages. Domain knowledge pertaining to radiology and brain hemorrhages such as “windowing” is applied for feature engineering to improve our target features for the CNNs.
Overall, we achieved a mean 90% AUROC in detecting hemorrhages depending on type with single slice CTs and consumer grade hardware translating to limited computational speed and memory.