Large Language and Generative Foundation Models for Cloud-Enabled Medical Imaging and Internet of Medical Things: A PRISMA Systematic Review of Architectures, Security, and Deployment

Abstract Large Language Models (LLMs) are becoming increasingly entwined in the cloud-based processing of medical images, enabling functions such as cross-modal synthesis, text generation, and decision support systems. In some of the earliest studies, the use of generative adversarial models (GANs) demonstrated the potential for MRI-to-CT transfer, and the synthesis of T1 and T2 images, and the ability to translate images without supervision using models such as CycleGAN. More recently, the approaches include the use of diffusion models, and the combination of GAN models and diffusion models, and multimodal models that enable the processing of imaging and clinical text simultaneously. Despite the progress, the key challenges include maintaining the consistency of the volumetric images for three-dimensional models, domain transfer between different scanners, and the need for models to enable reasoning for clinical interpretation. Using the systematic review approach based on the guidelines of the PRISMA protocol, the study reviews the literature regarding the differences and similarities between lightweight and complex models for the translation of images in the cloud-based and Internet of Medical Things (IoMT) systems, and the suitability of the models for implementation. Future studies will include the design of two-dimensional and three-dimensional models, and the development of models for distributed processing, which enable models to be scaled for different institutions for clinical decision-making. The study results provide conclusive insights into the fundamental trade-off between the two models, and the pivotal part of the new generation of models based on the design of the LLM systems.