A Nested Vision Transformer in Vision Transformer Framework for Intracranial Hemorrhage

This research aims to address challenges in the integration of graph neural networks with sequential data, particularly within the realm of medical limited images. To this end, we present a novel model called Vision Transformer in Vision Transformer (named ViV), which segments images into multiple patches and employs Transformer blocks for feature representation. This approach enables the comprehensive integration of global information and fine-grained details. Furthermore, to enhance the model’s stability when dealing with small batch sizes, we introduce RNSNorm with designing a TylerLoss loss function to better evaluate the model’s performance. Through the ViV model, we not only achieve significant advantages in both global and detail features but also excel in tasks such as predicting cerebral hemorrhage volume, thanks to the introduction of enhanced gradient propagation and convergence functions. On the RSNA-IHD dataset, the ViV model improves accuracy by more than 10% compared to other models. These achievements provide robust support for deep learning research in medical image analysis and various other domains. By effectively integrating graph neural networks with sequential data, the ViV model can not only enhance medical image analysis but also have broad applications in other fields.