Limited-supervision for efficient medical image understanding

动机：深度学习模型正在推动大多数视觉识别任务的进展，在医疗保健、自动驾驶或安全等战略领域具有巨大而宝贵的潜力。虽然这些技术在完全监督的情况下取得了出色的表现，但在实际应用中部署这些模型存在两个主要障碍：在新任务上的泛化性能差，以及在连续任务上训练时可能会丢失所学知识。理想情况下，将训练好的模型扩展到新的类别需要收集和标记新的类别的额外数据，并在整个新的数据集上重新开始训练过程。然而，这种设想是不现实的。因此，有必要将当前的学习算法提升到一个全新的水平，使模型能够快速适应新的任务（即，使用很少的标记样本）并避免知识遗忘。

Motivation: Deep learning models are driving progress in most visual recognition tasks, having an enormous and valuable potential in strategic domains such as health-care, autonomous driving or security. While these techniques have achieved outstanding performance when full-supervision is available, there exist two major obstacles to deploy these models in practical applications: poor generalization performance on new tasks and potential loss of learned knowledge when training on sequential tasks. Ideally, expanding a trained model to novel classes would require collecting and labeling additional data for new categories and restart the training procedure on the entire novel dataset. Nevertheless, this scenario is unrealistic. Thus, there is a need to bring current learning algorithms to a whole new level where models can quickly adapt to new tasks (i.e., with few labeled samples) and avoid knowledge forgetting. Research objectives and methodology: This research program will contribute to develop novel learning strategies for visual applications that address these two key challenges on deep models, with a primary focus on semantic segmentation. We intend to leverage unlabeled image data and prior domain knowledge, which have been underestimated in current approaches. This will enhance the representation power of deep models, leading to more generalizable features when few labeled data is available. To reduce the catastrophic interference between tasks -learned sequentially-, we will investigate novel attention mechanisms and the use of meta-memories, forcing the network to focus on class-agnostic parameters which will be retained across tasks. This research program will focus on medical images, given the unique challenges they bring to machine learning methods (e.g., complex shape, high variability, etc), and their tremendous potential for healthcare. Impact: I expect that this program will lead to robust learning methods to understand and interpret the content of medical imaging data in different challenging scenarios, such as low labeled data regime and sequential training, e.g., in decentralized systems. In the long-term, I aim at transferring specific human behaviour to computers, from an image interpretation perspective, so that they can efficiently learn from few samples and avoid significant memory losses when learning different tasks sequentially. Robust interpretation of these images will provide invaluable support for diagnosis, treatment and follow-up of many diseases. Even though devised methods can be applied to a breadth of applications, we will prioritize neurology, cardiology and oncology, which have a high economical and social impact and for which we have a considerable experience. Beyond the application domain adopted in this program, these novel learning strategies will make an impact on the area of semantic segmentation, particularly when full annotations are scarce, in a broad span of disciplines.