Deep Learning Methods for Detecting and Classifying Brain Hemorrhages with Different Preprocessing Approach

Intracranial hemorrhage (ICH), resulting from trau-matic brain injuries, poses significant risks if not promptly diagnosed. The current standard involves manual CT scan examinations by radiologists, which can be time-consuming and prone to errors. To improve this, deep learning techniques are being explored for ICH detection. Recent AI advancements show promise in enhancing the accuracy and efficiency of diagnosing ICH and its subtypes. By utilizing large datasets and advanced algorithms, AI-based systems can aid in the timely identifi-cation and localization of hemorrhages, potentially improving patient outcomes through quicker interventions. In this work, we implemented multiple models such as ResNet18 (2D and 3D), ResNet50, VGG16 and the Vision Transformer (ViT) architecture for classifying CT scans into Normal and Hemorrhage categories in a patient-wise manner. We utilized the MosMedData dataset, which includes 800 brain CT scans consisting of multiple series of DICOM images with and without signs of ICH. A significant challenge is that each patient has a different number of CT scan images, whereas our models require consistent depth across images. However, an advantage is that the CT scan images are sequentially ordered axially, allowing us to average pixel values and merge them into 3 and 10 channels. ResNet50 achieved the highest performance among our models using the 3-channel method, correctly identifying 278 out of 365 patients with brain hemorrhages and 270 out of 420 patients without hemorrhages.