Publication

Abstract : Convolutional neural networks have presented a huge breakthrough in automated medical diagnoses, specifically for Intracranial Hemorrhage (ICH) detection. However, it has been observed that most of them are either for detection in terms of identification or localization; trained on thousands of images. This work is a development for the complete and automated diagnosis of ICH using a smaller dataset, trained on unenhanced images using End to End Cascaded CNN. This includes multi-label classification of brain scan into one of its five subtypes, intraparenchymal, intraventricular, epidural, subdural and subarachnoid as well as the precise localization of the brain scan to localize the bleeding point. The VGG16 algorithm, used for multi-label classification 70.29% when trained on only 146 images. The U-net algorithm, used for semantic segmentation achieved an accuracy of 99.87%. Learning curves and hyper-parameter tuning have been used to an existing open sourced and smaller dataset to make deep learning models perform satisfactorily to address limitations of resource crunch in terms of systemic specifications as well as datasets.