Back to Projects List
Deep learning model for B-line detection in lung ultrasound videos using crowdsourced labels
Presenter location: In-person
- Mike Jin (Brigham and Women's Hospital, USA)
- Tamas Ungi (Queen's University, Canada)
- Colton Barr (Queen's University, Canada / Brigham and Women's Hospital, USA)
- Ameneh Asgari-Targhi (Brigham and Women's Hospital, USA)
- Tina Kapur (Brigham and Women's Hospital, USA)
Automated B-line detection in lung ultrasound videos has been demonstrated before, most recently by Lucassen 2023. However, acquiring the many labels necessary can be a resource-intensive process, limited by the availability of expert clinicians capable of producing high-quality labels. Recently, gamified crowdsourcing with a new quality control mechanism and built-in learning for labelers has been demonstrated to be capable of producing annotations on lung ultrasound videos comparable in quality to expert clinicians (as well as analogous results for EEG and skin lesion classification tasks), which can greatly shorten the time required to acquire high-quality labels for model training. Though these crowd labels have been shown to have expert-level quality, it has yet to be demonstrated whether crowd-produced labels are capable of training high-performance models.
- Train a deep learning model to classify lung ultrasound videos as having B-lines or having no B-lines.
Approach and Plan
- Create a data file associating all 3000+ clips with filepath, crowd label, and expert labels (for those that have expert labels).
- Adapt the model (ResNet(2+1)D-18 or similar pretrained model) and training procedure used in Lucassen 2023 to train a new model on a new crowd-labeled dataset of 3000+ lung ultrasound videos from 500 patients.
- Evaluate the model performance and compare to previously reported model performance for ultrasound video classification of B-line presence.
Progress and Next Steps
- De-identified and masked 3000+ lung ultrasound clips
- Uploaded 3000+ clips with standard filename format to a GPU cluster.
- Crowd-labeled all 3000+ lung ultrasound clips using 193 clips from ~70 patients for crowd training.
** PW39 progress **
- Lots of helpful discussions about model selection and handling varying input size
- Tried two different existing CNN + RNN solutions, but thwarted by hardware/environment setup/version compatibility issues.
- First time seeing functionality of 3D Slicer in greater depth, and was able to demo DiagnosUs annotation collection platform to some folks.
Background and References