Hybrid CNN-ViT Model for Intracranial Hemorrhage Detection and Classification in DICOM Images

Intracranial hemorrhage (ICH) refers to bleeding within the brain, a global concern that underscores the im-portance of early detection. ICH is typically detected using computed tomography (CT), which produces high-resolution scans commonly stored in the Digital Imaging and Commu-nications in Medicine (DICOM) format. In this paper, we propose three hybrid models combining Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) for ICH detection and subtype classification. The ViT model utilizes features extracted by different CNN variants for classification. Specifically, MobileNetV2, DenseNet121, and ResNet50 function as feature extractors for the ViT classifier, which incorporates a probabilistic residual multi-head self-attention mechanism to enhance focus on hemorrhage regions and improve classification performance despite noise and variability in CT images. We utilize the publicly available MosMedData dataset to illustrate the performance of our models. We have achieved an accuracy of 98.83% for binary classification using the proposed MobileNetV2-ViT model and an impressive 99.06% accuracy for subtype classification of epidural hemorrhage.