Artificial Intelligence


AI will play an increasingly important role in imaging research and its clinical application. Our Department focuses on the use of AI for both acquisition and post-processing in medical imaging, with the aim of developing new technologies and optimizing current techniques to improve healthcare through interdisciplinary research. We have ready access to large imaging databases for AI research. Additionally, our Department is equipped with several high-end computational power GPU Clusters to facilitate AI projects. To further support AI initiatives, we have established the Departmental CUHK lab of AI in Radiology (CLAIR) or AI Laboratory. 

Our current research interests in AI encompass a wide range of areas. By actively pursuing these research interests, we will positively contribute to the advancement of AI in medical imaging and improve patient care.

Deep learning for disease-specific applications

  • Nasopharyngeal carcinoma – Automated segmentation and detection of nasopharyngeal carcinoma (NPC) on MRI. We investigated the use of CNN to detect and segment primary NPC on plain MRI and compared the performance of deep learning in segmenting primary NPC on plain and contrast-enhanced MRI. We also developed AI algorithms that combine CNN and RNN to discriminate between early NPC and both benign hyperplasia, whose appearances can overlap, on plain MRI scans. 
  • Alzheimer’s disease – Automated brain segmentation for early detection of Alzheimer’s disease on MRI. 
  • Osteoporotic vertebral fractures – Automated opportunistic detection of vertebral fractures, specifically osteoporotic vertebral fractures.
  • Rheumatoid arthritis – Automated quantification of bone marrow edema and synovitis in rheumatoid arthritis on MRI. 
  • Knee osteoarthritis – Automated segmentation of knee tissues, quantification of cartilage thickness, and characterization of knee osteoarthritis.  
  • Multiple sclerosis and microbleeding – We are developing AI models for automated brain segmentation in neurodegenerative disorders and multiple sclerosis, AI-based detection of intracerebral hemorrhage and mass effect on head CT exams, and AI assessment of acute ischemic stroke.
  • Maxillary sinus lesions – We developed and validated an AI model to segment lesions in the maxillary sinus on cone-beam CT that discriminates whether the lesion is a focal cyst or diffuse hyperplasia. Accurate characterization is crucial to the outcome of dental implant surgery.

Radiomics research

  • Lesion characterization – Apart from deep learning methods, we have also investigated radiomics that are better at characterizing nasopharynx lesions and discriminating them from early nasopharyngeal carcinoma. 
  • Deep-learning-based radiomics – Segmentation of lesions on images using deep learning AI is much more established and foolproof when compared to direct classification using deep learning. We have, therefore, combined the interpretable classifier built using radiomics with an accurate automatic segmentation algorithm for maxillary sinus lesions to characterize and discriminate between cysts and hyperplasia, which are often difficult to distinguish on CT.
  • Robustness and reliability of radiomics pipeline – We investigated and devised an ensemble method for feature selection in radiomic pipeline that improved the feature selection stability and, hence, model explainability. We also showed that fine-tuning only the feature normalization parameters can partially recover the performance of a trained model when dealing with data from untrained, but related, knowledge domains, thus improving the robustness of the radiomic pipeline.

Advanced Imaging through Deep Learning

  • Advanced sequence development – We are exploring non-invasive biochemical MRI for early diagnosis and treatment monitoring with automated post-processing and quantification. This involves using AI to develop MRI denoising methods, physics-informed quantification methods, tissue segmentation, tissue phenotype analysis, uncertainty analysis of quantification, and automatic ROI selections. 
  • Fast MRI reconstruction – Deep learning methods for rapid anatomical MRI and quantitative MRI. Our multi-prior deep learning approach for highly accelerated MRI received a third Place Award at CMRxRecon Challenge organized by STACOM Workshop in MICCAI 2023, Vancouver, Canada.
  • Brain mapping – We are developing a graph convolutional network to improve our understanding on brain function. Additionally, we are working on an AI model for Brain radiology report generation.
  • Brain characterization – We have developed fast deep learning quantification methods for in vivo MRI fingerprinting, and a deep learning pipeline for quantitative susceptibility mapping.
  • Chest X-ray report generation – We are developing a foundation model for generating annotated radiology reports for different chest X-ray applications.

Translational research

  • Screening – We have developed an AI solution that aims to screen for early nasopharyngeal carcinoma using a plain MRI sequence, which is quick, radiation-free, non-invasive, requires no contrast agent, and can automatically generate a comprehensive screening report without the need for a radiologist. This solution aims to run at a low-cost as a cost-effective screening program.  We are prospectively validating the effectiveness of this solution, which will be crucial for clinical translation.
  • Interpretability – We have developed a slice-wise attention mechanism for CNN that automatically learns to take into account the varying importance of different slices of an image dataset when conducting a diagnostic task. We are also actively researching other means to unveil the inner workings of the notorious deep learning “black box” that is currently a major challenge to translation of AI research to clinical use, especially with 3D imaging modalities. 
  • Generalizability – We are developing methods to address domain shift problems among datasets collected from different protocols, software, and hardware, which is an issue commonly encountered in clinical practice.
  • Patient privacy – Our Department is actively developing techniques to address challenges from medical imaging data privacy issues for developing AI techniques and digital data transfer.