Breast Cancer
Conditions
Keywords
mammography, artificial intelligence, deep learning
Brief summary
Breast cancer (BC) is the most common cancer among women in worldwide and the second leading cause of cancer-related death. As the corner stone of BC screening, mammography is recognized as one of useful imaging modalities to reduce BC mortality, by virtue of early detection of BC. However, mammography interpretation is inherently subjective assessment, and prone to overdiagnosis. In recent years, artificial intelligence (AI)-Computer Aided Diagnosis (CAD) systems, characterized by embedded deep-learning algorithms, have entered into the field of BC screening as an aid for radiologist, with purpose to optimize conventional CAD system with weakness of hand-crafted features extraction. For now, stand-alone performance of novel AI-CAD tools have demonstrated promising accuracy and efficiency in BC diagnosis, largely attributed to utilization of convolution neural network(CNNs), and some of them have already achieved radiologist-like level. On the other hand, radiologists' performance on BC screening has shown to be enhanced, by leveraging AI-CAD system as decision support tool. As increasing implementation of commercial AI-CAD system, robust evaluation of its usefulness and cost-effectiveness in clinical circumstances should be undertaken in scenarios mimicking real life before broad adoption, like other emerging and promising technologies. This requires to validate AI-CAD systems in BC screening on multiple, diverse and representative datasets and also to estimate the interface between reader and system. This proposed study seeks to investigate the breast cancer diagnostic performance of AI-CAD system used for reading mammograms. In this work, we will employ a commercially available AI-CAD tool based on deep-learning algorithms (IBM Watson Imaging AI Solution) to identify and characterize the suspicious breast lesions on mammograms. The potential cancer lesions can be labeled and their mammographic features and malignancy probability will be automatically reported. After AI post-processing, we shall further carry out statistical analysis to determine the accuracy of AI-CAD system for BC risk prediction.
Interventions
OTHERmammography
standard mammography including craniocaudal (CC) and mediolateral oblique (MLO) views
Sponsors
IBM China/Hong Kong Limited
Chinese University of Hong Kong
Study design
Observational model
COHORT
Time perspective
RETROSPECTIVE
Eligibility
Sex/Gender
FEMALE
Healthy volunteers
Yes
Inclusion criteria
* Women who had undergone standard mammography including craniocaudal (CC) and mediolateral oblique (MLO) views.. * Histopathology-proven diagnosis is available for patients with breast malignancy, including invasive breast cancer, carcinoma in situ, and borderline lesion et al. * As reference standard of benign nature, results from pathology or clinical long-term follow-up (\>=2 years) examinations are available for cases without breast malignancy.
Exclusion criteria
* Patients with concurring lesions on mammograms that may influence subsequent AI post-process. * Patients without available pathologic diagnosis or long-term follow-up (\>=2 years) examinations. * Patients who had undergone breast surgical intervention (e.g. lumpectomy and mammoplasty) prior to first mammography. * Patients diagnosed with other kinds of malignancy, concurrent with metastasis or infiltration/invasion to breast.
Design outcomes
Primary
| Measure | Time frame | Description |
|---|---|---|
| sensitivity | 3 years | true positive rate in percentage(%) derived by ROC analysis |
| specificity | 3 years | true negative rate in percentage (%) derived by ROC analysis |
| area under curve (AUC) | 3 years | area under receiver operating characteristic (ROC) curve in percentage (%) |
| accuracy | 3 years | proportion of true results(both true positives and true negatives) among whole instances |
Countries
Hong Kong
Outcome results
None listed