Neusoft Research of Intelligent Healthcare Technology

Collaborative Lab of Health Informatics with Neusoft

Stream 1. Text Processing and Knowledge Graph Development

With the increasing demand for clinical decision support systems and self-diagnostic symptom checkers, considerable effort has been devoted to constructing health knowledge graphs on which these diagnostic platforms rely. Most of existing health KGs are manually compiled by experts through a labour-intensive process, making them extremely brittle and difficult to adapt to new clinical settings. This project will explore an automated process to learn high quality KGs from various medical data including textbooks, journals, trusted web content, and electronic medical records. Natural language processing and statistical analysis will be applied to extract medical entities such as symptoms, diseases, drugs, etc. from unstructured and semi-structured data, as well as correlation and causation between these entities. We regard it as the first step towards developing learning models that perform diagnostic inference directly on top of the automatically built health knowledge graphs.

Stream 2. Medical Image Analysis

This project will develop a unified framework to automate/help manual diagnosis, and optimize workflows of treatments by taking advantage of recent developments and advancements in medical imaging research. More specifically, we will address the biggest research challenges in medical imaging: 1) improving the quality and speed of medical image annotations; 2) building medical imaging datasets using active learning; 3) classifying disease and conditions by accurately detecting and localizing regions of interest in medical images using a multi-layer CNN approach; and 4) providing accurate segmentation and localizing the boundary of objects in medical images. An example application is reducing the risks of radiation therapy by providing accurate segmentation of the boundary of health tissues and tumour. We will also focus on saving the time of patients by converting or transferring knowledge of different imaging procedures (i.e. converting MRI images into CT images, constructing a routine CT from a low dose CT). We will build upon our expertise in transfer learning, image annotation, and recent advances in deep learning and more specifically in generative adversarial networks.

Stream 3. Human-in-loop Quality Management

Complementary to other two projects, we will design human-in-the-loop techniques to support and improve the results of other methods developed across all stages of the project. For example, we will look at 1) improving the use of manual annotations to support AI methods for medical image processing; 2) the use of human feedback to improve the quality of automatic entity relationship extraction and knowledge graph construction; and 3) the use of human judgements to train and evaluate medical information retrieval models. This part of the project will specifically focus on medical data annotation tasks. For this we will design novel task routing algorithms to assign a manual annotation task to the right expert considering at the same time 1) the cost to access experts, 2) the skills required to complete the task, and 3) the task priority. We will also be looking at tasks design strategies that can minimise the time required by experts to complete the task effectively.

Chief Investigator(s): Dr Guido Zuccon (UQ), Dr Sen Wang (UQ), Dr Gianluca Demartini (UQ), Dr Mahsa Baktashmotlagh (UQ), Dr Wen Hua (UQ)

Administering Organisation: UQ

Value: $1,058,252.00 (AUD)

Founding round: 2019 (2019-2022)