| 研究実績の概要 |
As most of the existing automatic surgical video analysis models require a large number of manually labeled data for training, this project aims to design a learning method to perform spatial and temporal segmentations with smaller requirements of humans’ input. During this project, I mainly studied the following sub-topics towards the goal.
1.Surgical images/frames analysis using very few training samples. I developed an explainable few-shot learning method to give accurate recognition labels (as well as the explanations) to the input samples, which is very important for risk-sensitive areas like medicine. This work is presented at CVPRW 2021.
2.Surgical images/frames semantic segmentation in a weakly-supervised way. I developed a new training strategy for video semantic segmentation models to utilized unlabeled data to improve their segmentation performance. This work is published in IEEE Access.
3. Computer vision models that can output prediction results as well as visual explanations for not only medical images but also natural images. The explanations can help downstream tasks like semantic segmentation, etc. Therefore, it has the potential to enable the weakly-supervised surgical images/frames semantic segmentation when only frame-level labels are available. This work is presented at IEEE ICCV.
4.Surgical videos temporal analysis using no labels. I developed a retrieval-based method to automatically predict surgical duration. This work is under submission.
|
