Generative Adversarial Approach for Medical Image Synthesis with Context Aware Deep Convolutional Networks
Vaibhav Bhogade, Nithya B S · 2024
Medical imaging plays a crucial role in various clinical applications. However, due to factors like cost and radiation dose, acquiring certain image modalities may be limited. Therefore, medical image synthesis can be highly advantageous as it estimates desired imaging modalities without the need for actual scans. In this paper, a generative adversarial approach is used to address this challenging problem. To create a target image from a source image, a fully convolutional network (FCN) is specifically trained. An adversarial learning approach is used to improve the FCN’s capacity to simulate the nonlinear mapping from source to target and generate more accurate target images. Additionally, an image-gradient-difference based loss function is added into the architecture of the FCN to prevent hazy target images. Investigating long-term residual units helps to enhance the network’s training procedure. An Auto-Context Model is also used to develop a context-aware deep convolutional adversarial network. The correctness and reliability of the proposed method for creating target images from corresponding source photos are shown by experimental findings. The proposed methodology is assessed using datasets for the tasks of producing CT pictures from MRI and producing 7T MRI images from 3T MRI data. The results will demonstrate how well this technique handles picture generation for a variety of medical imaging tasks