Medical Artificial Intelligence, Medical Imaging
Conditions
Brief summary
Biomedical deep learning (DL) often relies heavily on generating reliable labels for large-scale data and highly technical requirements for model training. To efficiently develop DL models, we established an integrated platform to introduce automation to both annotation and model training-the primary process of DL model development. Based on this platform, we quantitively validated and compared the annotation strategy and AI model development with the pure manual annotation method performed on medical image datasets from multiple disciplines.
Interventions
None listed
Sponsors
Sun Yat-sen University
Study design
Observational model
CASE_CONTROL
Time perspective
CROSS_SECTIONAL
Eligibility
Sex/Gender
ALL
Healthy volunteers
Yes
Inclusion criteria
* have medical imaging record (including ophthalmology, pathology, radiography, blood cells, and endoscopy)
Exclusion criteria
* unqualified medical imaging
Design outcomes
Primary
| Measure | Time frame | Description |
|---|---|---|
| annotation accuracy | baseline | calculate annotation accuracy for comparison between groups with using the annotation results |
Secondary
| Measure | Time frame | Description |
|---|---|---|
| accuracy of model performance | baseline | calculate AI model accuracy for comparison between groups with using the model predicted results |
| AUC of model performance | baseline | calculate AI model AUCs for comparison between groups with using the model predicted results |
| annotation time cost | baseline | calculate annotation time cost for comparison between groups with using the time recorded during the tests |
Countries
China
Contacts
Primary ContactHaotian Lin, Ph.D, M.D.
Outcome results
None listed