Multimodal Medical Image Registration Classification Framework Using Squeeze-and-Excitation Network

Medical image registration is an essential procedure in medical image analysis. Traditional 2D and 3D registration techniques often face challenges in addressing the intricate variations and conditions found in medical images. Complex deformation issues are commonly encountered in these registration tasks. Although several advanced registration models are developed for registering the medical images. Yet, achieving accurate and efficient deformable registration remains a significant challenge, particularly when dealing with substantial volumetric deformations. Given the difficulties in obtaining ground-truth registration labels, many existing methods utilize an unsupervised multi-modal image registration approach. However, designing effective metrics to assess the similarity of multi-modal images is challenging, which significantly hampers the performance of multi-modal registration. Additionally, due to the differences in contrast various modalities are utilized for collecting the raw images and it becomes a difficult task. Therefore, addressing the numerous challenges associated with traditional multi-modal medical image registration techniques, a novel deep learning approach is designed. Initially, multi-modal image registration and segmentation procedures are performed in this research work. Here, the image registration and segmentation procedure are executed using a UNet-based Autoencoder with Vision Transformer (UNet-A-ViT). Next, the registered and segmented images are provided to the classification phase. In this phase, image classification procedures are performed using Squeeze-and-Excitation Network (SENet) and attain the classified outcomes like malignant and benign tumors. Finally, various experiments are executed in the suggested framework to observe its efficiency over classical techniques.