Automated classification and segmentation of Intracranial Hemorrhages using 2D Convolutional Neural Networks and U-Net architecture on Computed Tomography scans

Intracranial hemorrhage (ICH) is a type of stroke and a life-threatening condition that requires rapid and accurate diagnosis. This research utilized advanced Convolutional Neural Networks (CNNs) and a U-Net architecture to automate the classification and segmentation of ICH in Computed Tomography (CT) scans to improve the accuracy and efficiency of the diagnosis of ICH. Three primary models were constructed, 1) a 2D binary CNN model for classifying individual slices of CT scans as being hemorrhagic or non-hemorrhagic, 2) a 2D multi-label CNN model for diagnosing the type of hemorrhage for only hemorrhagic scans, and 3) a 2D U-Net model for segmenting hemorrhagic regions within the individual CT slices already classified as hemorrhagic from the 2D binary classification model. The results showed that the binary classification model, the multi-label classification model and the U-Net segmentation model achieved test accuracies of 97.7%, 81.3% and 98.7% respectively. These findings highlight the potential of CNNs and image segmentation to assist radiologists in faster and more accurate ICH detection and diagnosis.