A Hybrid ResNet and ViT Model for Domain Generalization in Chromosome Classification

Human chromosome identification is an important part of karyotype research and plays a vital role in birth defect diagnosis and biomedical research. Currently, various computer vision-based classification methods have achieved good results in chromosome classification tasks. However, because chromosomes taken by different hospitals using different electron microscopy equipment have slightly different characteristics, the performance of existing models decreases when generalizing to new chromosome data. To address this issue, we propose a model named RVC that integrates ResNet and Vision Transformer architectures to extract multi-scale features from images, thereby enhancing the network's capability to capture both global and local information. At the same time, we use the cosine similarity loss function to learn intra-class commonalities and inter-class differences to further improve the model's generalization ability. Furthermore, a specific feature normalization module is introduced for our model to achieve reliable chromosome classification generalization task. Our method is validated on the chromosome dataset. A large number of experimental results show that this method can significantly improve the generalization performance of the model.