- Abstract: Machine learning applications in medical imaging are frequently limited by the lack of quality labeled data. While conventional active learning approaches have been able to reduce the labeling burden to some extent, the main difficulty was defining an effective sampling criteria. In this work, we propose a novel framework, semi-supervised reinforced active learning, which utilizes inverse reinforcement learning and an actor critic network to train a reward based active learning algorithm. This is an extension of the reinforced active learning formulation to complex problems where direct rewards may be unavailable. The framework was tested on a U-Net segmentation network for pulmonary nodules in chest X-rays. The proposed framework was able to achieve the same level of performance as the standard U-Net while using only 50% of the labeled data, demonstrating ability to effectively reduce the labeling burden.
- Keywords: Chest X-ray, Pulmonary nodules, Semi-supervised learning, Reinforcement learning, Reinforced Active Learning, Semantic segmentation
- Author Affiliation: Vuno Inc., Imperial College London