Computer-Aided Diagnosis of Pulmonary and Neurological Disorders
Lead researchers: A/Prof Suhuai Luo, Dr Raymond Chiong, Dr Peter Summons, Mr Amir Ebrahimi-Ghahnavieh, Mr Liton Devnath, Mr Rehan Afzal
Partners: Dr Ron Li, Dr Dadong Wang, Dr Xuechen Li, Prof Linlin Shen, Prof Jeannette Lechner-Scott, A/Prof Saadallah Ramadan
A computer-aided diagnosis (CAD) system can help doctors, radiologists and other medical practitioners make better clinical decisions. The success of CAD depends on the successful integration of advanced signal processing, image processing and artificial intelligence and machine learning techniques, as well as a large amount of training data.
In CAD, deep neural networks ‘learn’ to perform tasks by considering known data inputs and outputs. At the University of Newcastle, a team of researchers are focused on taking original medical images as inputs, applying pattern recognition algorithms to perform diagnosis and comparing results to the work of multiple experts. By repeating this process, the network learns to become more accurate and adaptively improves its performance to achieve ‘superhuman’ results; significantly more precise than a typical result through a manual process.
The team are particularly focused on applying CAD to neurological disorders such as Alzheimer’s disease, multiple sclerosis, brain tumor, Parkinson’s disease, epilepsy, stroke, as well as pulmonary diseases like asthma, chronic bronchitis, emphysema, infections (such as influenza and pneumonia), lung cancer, and pulmonary fibrosis.
To enhance their results, the team are applying sophisticated methods such as:
- Long short-term memory techniques that process entire sequences of data, for example – sets of medical images that show the progression of a disease over years, to help make better predictions about prognosis.
- 3D convolutional neural networks, that extract features from both the spatial and temporal dimensions of the data set.
- Data augmentation processes, such as Cycle-GAN and the Keras Image Data Generator, that artificially extend the diversity of data available for training models.
- Transfer learning, where knowledge gained from one training exercise is applied to a new but related problem.
- Generative adversarial networks (GANs) in which two neural networks compete against each other to synthesise new data.
Specific projects include: ‘Deep learning algorithm for detection and prognosis of Alzheimer’s Disease’; ‘Predicting disability progression in multiple sclerosis patients for optimal treatment and management’; ‘An accurate black lung detection using transfer learning based on deep neural networks’; and ‘A Solitary Feature-based Lung Nodule Detection Approach for Chest X-Ray Radiographs’.