Unsupervised, pre-trained models for image quality and workflow improvement in medical application

Abstract

Recently, deep learning techniques have advanced significantly contributing to notable progress in the medical field. In particular, the unsupervised method is valuable in the medical field where acquiring labeled datasets can be challenging. Another notable approach involves the pre-trained model, which was trained on a large-scale dataset encompassing both vision and text data. This model achieved contextual and cross-modal understanding, showing significant performance improvements in various tasks. this study introduces a method aimed at addressing various challenges related to image quality degradation in the medical imaging field and enhancing workflow. I introduced a method for translating 3-dimensional (3-D) ultrasound image quality to 2-D quality, enhancing contrast, sharpness, and reducing artifacts. Additionally, I proposed a method for RealisticVue image restoration and enhancement network, aiming to generate a realistic full facial view of the fetus for a more pleasing experience for pregnant. Furthermore, I suggested a method for ultrasound image translation using speech guidance to achieve image control without direct handling. Moreover, I presented a method for text correction in the medical domain, generating from conventional Speech-To-Text (STT) systems by leveraging vision information. Finally, I proposed a method for breast ultrasound report generation using an LLM-based LangChain framework, addressing the time-consuming issues. I demonstrate that my approach holds considerable promise for practical implementation in the clinical field.