Data Collaboration Lab

Project Summary:

Endovascular mechanical thrombectomy is now an established treatment for acute ischemic stroke caused by large vessel occlusion (LVO). As the majority of ischaemic stroke patients would first undergo plain Computed Tomography (CT) brain scan, it is imperative that signs of LVO on plain CT scans can be identified quickly and reliably so to initiate confirmatory investigations, clinical referral and proper treatment. The objective of the study is to develop a novel, rapid and automated computer algorithm capable of detecting and predicting signs and likelihood of LVO. The ultimate goal is to generate an effective, reliable and locally relevant platform for triaging acute stroke patients.

 

Publication:

1. You, Jia, Yu, Philip L.H, Tsang, Anderson C.O, Tsui, Eva L.H, Woo, Pauline P.S, Lui, Carrie S.M, Leung, Gilberto K.K, Mahboobani, Neeraj, Chu, Chi-yeung, Chong, Wing-ho, Poon, Wai-lun (2021). 3D dissimilar-siamese-u-net for hyperdense Middle cerebral artery sign segmentation. Computerized Medical Imaging and Graphics, 90, 101898. (Link)
2. You, Jia, Tsang, Anderson C O, Yu, Philip L H, Tsui, Eva L H, Woo, Pauline P S, Lui, Carrie S M, & Leung, Gilberto K K. (2020). Automated Hierarchy Evaluation System of Large Vessel Occlusion in Acute Ischemia Stroke. Frontiers in Neuroinformatics, 14, 13. (Link)

 

 

▼

Project Summary:

A comprehensive review of the secular trend and current situation of chronic viral hepatitis would provide pivotal data to the Government and the Steering Committee to formulate strategies to effectively prevent and control viral hepatitis to guide the strategies to achieve the goals set by the World Health Organization. This project reports the increasing uptake of antiviral treatment in patients with chronic viral hepatitis from 2000 to 2018. The researchers are developing a machine learning model as the risk stratification tool and to predict important clinical outcomes. A machine-learning model based on clinical parameters available in the Clinical Management System (CMS) are being developed to facilitate the clinical assessment and risk stratification of liver complications including liver cancer in patients with chronic viral hepatitis. The machine learning computer-aided diagnosis for liver nodules are being optimised. The findings will provide important data to address the request from the Chief Executive's 2017 Policy Address to formulate strategies to effectively prevent and control viral hepatitis, in particular for the planning of upcoming strategies to eradicate chronic viral hepatitis by year 2030 according to the advocate of World Health Organization.

 

Publication:

1. Wong GL, Tse YK, Yuen BW, Yip TC, Wong VW, Chan HL, Yuen PC. Secular trend of treatment uptake in patients with chronic viral hepatitis: a territory-wide study of 120,279 patients with data from HADCL from year 2000 to 2017. Hepatol Int 2020;14(Suppl 1):A637.  (Link)
2. Wong GL, Tan Q, Tse YK, Yang BY, Yip TC, Yin C, Hui VW, Yuen BW, Chan HL, Wong VW, Yuen PC. Machine learning models to predict hepatocellular carcinoma in patients with chronic viral hepatitis – a territory-wide study from hospital authority data collaboration lab (HADCL) in 2000-2017. Hepatology 2020;72(Supp 1):466A. (Link)
3. Wong GL, Yin C, Tse YK, Lyu F, Yip TC, Hui VW, Chan HL, Wong VW, Yuen PC. Deep learning of CT images in patients with chronic viral hepatitis – a territory-wide study with data from HADCL from year 2000 to 2017. Hepatol Int 2021;15(Suppl 1):A74. (p.123 of 306) (Link)

 

 

▼

Project Summary: Coming soon.

Publication: Coming soon.

 

 

▼

Project Summary:

Mortality and morbidity rates (MMR) are over 70% for patients admitted into emergency with a ruptured aneurysm. MMR is less than 7% if the patients are admitted before rupture. Currently, more than 200,000 people in Hong Kong are at risk of cerebral aneurysm rupture. Cerebral Aneurysm Risk Index (CARI) developed in this project can help. Early detection of cerebral aneurysm using CARI can reduce MMR to less than 7% and facilitate preventive treatment before rupture. Development of Phase I CARI is led by HKUST and KWH in collaboration with HADCL.

• CARI provides quantified aneurysm risk index to facilitate and support clinical decisions
• Helps identify high-risk patients for treatment before rupture
• Decision support to lower MMR of high-risk patients

 

Publication: Coming soon.

 

 

▼

Project Summary: Coming soon.

Publication: Coming soon.

 

 

▼

Project Summary:

Across HA hospitals, a high volume of brain CT scans are performed daily in the accident & emergency departments (AEDs) for the assessment of patients with acute conditions such as trauma, sudden headache, dizziness or stroke. One of the critical findings in CT scans is bleeding in the brain or “intracranial haemorrhage” (ICH). ICH requires timely identification as its presence and type determine the next step in the patient's clinical management. Conversely, its absence is equally important in the treatment of acute stroke and safe discharge of patients.

CT scans are 3D stacks of 20-300 images of the brain, and may be comprised of thick slices (~5mm) or thin slices (<1mm). Radiologists are specialised in interpreting CT scans but in practice, when patients undergo brain CT scans in the AED, initial interpretation of these images relies heavily on frontline emergency physicians, who may have limited experience and time in reading CT scans. Several studies have demonstrated that detection of ICH on brain CT scans using artificial intelligence (AI) can be accurate and useful for triage purposes, and for our purpose can potentially assist emergency physicians. In this project, we used a convolutional neural network - based multi-label classification task to develop a deep learning model for ICH detection. The model was first trained on an international publicly available dataset of over 670,000 expert-labeled CT slices, and independently tested on CT scans from the HA Data Collaboration Laboratory which constituted “unseen data”. The model output provided (1) an overall probability of ICH per CT scan and nominated (2) 5 potential ICH-positive CT slices for human cross reference.

The multi-center CT scans from HA Data Collaboration Laboratory (HADCL) were retrieved. The CT scans comprised of thin- and thick slices and were expert-labelled by radiologists for the presence or absence of ICH. Diagnostic accuracy of the model was determined using Area under the receiver operating characteristics curve (AUC) analysis; AUC ranges from 0 to 1 where 0 indicates a perfectly incorrect classification and 1 indicates perfectly correct classification. Since the model provided an overall probability of ICH, we used a probability threshold at approximately 80% sensitivity for binary classification of the model result into ICH-positive or ICH-negative. The corresponding specificity, positive predictive value (PPV) and negative predictive value (NPV) were then calculated accordingly. We assessed the diagnostic performance at the CT scan-level with subgroup analysis for both thin and thick slice CT scans. We additionally performed post-hoc labelling of the model-nominated CT slices to assess diagnostic performance at the slice-level.

Results for CT scans with thin slices

There were 2397 CT scans with thin slices. Of these, 1329 were ICH-negative and 1068 were ICH-positive. The model showed AUC of 0.81 in this cohort. An ICH probability threshold of 35% provided sensitivity of 81%, with specificity, PPV and NPV of 60%, 72% and 72%, respectively.

In 261 CT scans with post-hoc labelling of 5 nominated slices (total: 1305 CT slices), AUC was 0.85. A threshold of 45% provided sensitivity of 80%, with specificity, PPV and NPV of 74%, 87% and 63%, respectively.

Results for CT scans with thick slices

There were 389 CT scans with thick slices. Of these, 46 were ICH-negative and 343 were ICH-positive. The model showed AUC of 0.83 in this cohort. An ICH probability threshold of 50% provided sensitivity of 80%, with specificity, PPV and NPV of 67%, 95% and 31% respectively.

In 308 CT scans with post-hoc labelling of 5 nominated slices (total: 1540 CT slices), AUC was 0.88. A threshold of 60% provided sensitivity of 80%, with specificity, PPV and NPV of 83%, 91% and 65%, respectively.

Comparison of CT scans with thin and thick slices

There were 382 CT scans with both thin and thick slices. Of these, 342 were ICH-positive and 40 were ICH-negative. AUC’s were comparable (0.80 for thin slice CT scans vs. 0.82 for thick slice CT scans). At 80% sensitivity, the specificity/PPV/NPV were 53% / 93% / 23% for CT scans with thin slices and 65% / 95% / 27% for CT scans with thick slices.

Our results show that a deep learning tool allows detection of ICH in multi-center CT scans in the HADCL, therefore has potential to be used under the HA setting. Diagnostic accuracy appears comparable between thin and thick slice CT scans. However, for the intended purpose in the AED, further model improvement should be directed towards more accurate detection of ICH on CT scans with thick slices, and in facilitating cross checking of AI results by emergency physicians.

Publication: Coming soon.

 

 

▼