Accurate lesion segmentation in medical images is important to guide clinical management and surgical planning but is laborious when performed manually. Supervised deep learning methods have made remarkable progress in automating this process, leading to significant savings in time and manpower. However, applications based on such methods still require large-volume image datasets fully labelled at the pixel level to train artificial intelligence (AI) models.
Some studies mitigate this challenge with approaches using mixed supervision or semi-supervision. The underlying strategy is to find ways of requiring only a small fraction of data to be completely annotated at the pixel level while the remaining images contribute to "weaker" supervision by being annotated at the image level or even not at all. Sampling strategies developed for data selection coupled with active learning approaches also enable users to prioritize data labelling to maximize/enhance model performance under constraints on the annotation budget. Although these approaches address the data scarcity issue by leveraging both limited pixel-level labelled data and more readily-available image-level labelled/unlabelled data, they have not been translated to AI-based clinical imaging annotation platforms to demonstrate their effectiveness and impact in reducing annotation burden.
To address this gap, we propose an integrated and modularized annotation platform easily extended to support customized deep learning models (full supervision, mixed supervision, semi-supervision, etc.) and sampling approaches (random sampling, active learning, etc.), which enables researchers to perform rapid data annotation for training AI models with limited labelled datasets. Additionally, we engineer it to work with multiple third-party segmentation tools (e.g., ITK-Snap, 3D Slicer) so it can function as a unified and flexible platform that allows users to choose various viewing/sampling/learning strategies for different datasets and tasks. We will illustrate an instantiation of our proposed integrated and modularized annotation platform by describing an example use case of a segmentation task employing mixed supervision.
License type:
Publisher Copyright
Funding Info:
This research is supported by core funding from: Institute for Infocomm Research, Science and Engineering Research Council
Grant Reference no. :