Leveraging language understanding to decrease uncertainty measurements in medical image analysis

点击翻译按钮获取中文摘要

As per the UKRI portfolio and priorities, this project falls within the EPSRC Medical imaging -including medical image and vision computing research area.In the rapidly advancing field of medical image analysis, there is a growing need for robust and reliable tools to aid in clinical decision-making. The increasing complexity of imaging data, coupledwith the inherent uncertainty in medical diagnostics, outlines the significance of a better understanding of uncertainty measurements in medical image analysis. The research project aims to address this critical gap by delving into the realm of uncertainty in medical image analysis, withthe ultimate goal of improving the accuracy and reliability of diagnostic outcomes using additionallinguistic information.In general, modern advancements in medical image processing have derived from the use ofartificial intelligence (AI) which learns from a set of training data. Once the model has beentrained, it is then queried with new unseen data so that we can assess it's performance and evaluate deploying the model to use in the real-world. The term for this AI in the context of 2D and 3D images is called computer vision and encompasses how computers understand the semantics of imaging data.Rapidly growing advancements in computer vision applications have been leveraging powerfultools from other areas of deep learning such as from Natural Language Processing (NLP) which uses AI to understand and act on language information. As a result, training procedures and models are improving. For example, transformers which are a cutting edge methodology to understand language, utilise auxiliary problems when training. These are alternative sub problems that the model is trained on that have been proven to improve the performance on the original primary problem.Additional existing research can highlight anomalous regions of a scan without using labelledtraining data to say "this region is anomalous and contains a glioma". This is referred to asunsupervised learning and generally involves using a generative model such as a VAE, GAN ordiffusion model to recreate the image from a compressed representation of the image. Consequently this allows us to lean in to the limited anomalous data that may be available from a medical perspective.Aims1. Evaluate the robustness of existing uncertainty measurements in medical image analysis.2. Quantify the impact of uncertainty on diagnostic accuracy and clinical decision outcomes.3. Explore the application of uncertainty-aware models in diverse medical imaging modalities,such as MRI, CT, and X-ray.4. Investigate the consequence adding natural language has on the uncertainty on clinical diagnostic accuracy.