Weakly-supervised learning in Computer Vision and Medical Imaging

Abstract

Weakly-supervised learning is a fundamental problem in computer vision and medical imaging, aiming to learn from imperfect supervision signals. Deep neural networks have been the dominant model behind current solutions, achieving great success in various application domains. Weakly-supervised learning can be formulated as: 1) incomplete supervision, where only a small subset of training samples are annotated, 2) inaccurate supervision, where some training samples are annotated with incorrect supervision, and 3) inexact supervision, where training samples are given ambiguous or indirect supervision signals. Despite the remarkable achievements of current approaches, there are still many challenges worth exploring to advance the field, particularly in real-world datasets containing non-ideal scenarios. State-of-the-art incomplete supervision methods such as semi-supervised learning focus on consistency regularization and explore various data augmentation strategies. However, in real-world scenarios such as Medical Image Analysis (MIA), these methods fail with severe class imbalance. Moreover, few methods have been tested under a multi-label setup, which is common in MIA. Therefore, we argue that SSL methods in MIA need to be flexible enough to handle both multi-class and multi-label, as well as imbalanced learning. To address this problem, we propose two approaches that utilize self-supervised learning and pseudo-labelling to address the aforementioned issues and consequently improve semi-supervised learning performance on MIA tasks. For inaccurate supervision such as noisy label learning, the focus is mainly on balanced multi-class classification with sample selection methods. To solve noisy label learning in MIA, we propose to use a new regularization loss that considers both noisy labels and imbalanced learning for MIA. Furthermore, we utilize multi-modality information to better re-label multi-label images in MIA. Our results on MIA benchmarks show our state-of-the-art performance and effectiveness. We also understand noisy label learning from an inexact supervision perspective by learning from label sets instead of single label supervision for multi-class classification. We analyze the advantage of multi-label learning and partial label learning in noisy label learning and demonstrate the unique property of learning with label sets.