Enabling Autonomous Data Annotation in Mammography Image: A Human-in-the-Loop Reinforcement Learning Approach

The vast majority of work in computer vision focuses on developing and applying new machine learning models, which often rely on large amounts of labeled training data. However, annotation is costly and time-consuming. This paper presents an approach based on Deep Reinforcement Learning (DRL) to automatically generate new annotations and reduce human effort in the preparation of training data for supervised learning models in object detection. Our methodology introduces a virtual agent trained with human guidance, inspired by constructivist teaching methods, where interaction with a human teacher supports the agent's learning process. The proposed approach, named "Try a Little More" (TLM), employs active learning to identify uncertain cases and request human intervention during training, progressively enhancing the agent's ability to annotate autonomously. We evaluated our methodology on a mammography dataset, where the agent created bounding box annotations later used in a state-of-the-art supervised object detection algorithm. The results demonstrate that the proposed approach improves the quality and quantity of training data, surpassing existing reinforcement learning and human-computer interaction methods. Quantitatively, TLM generated 414 new annotations with an IoU of 0.86 and F1-score of 0.92, while increasing the mAP of a YOLO detector from 0.52 to 0.91, representing a 75% improvement in detection performance with 35% less human intervention. This contribution advances the field of Data-Centric AI by introducing a teaching-inspired methodology that combines human advice and reinforcement learning to accelerate the creation of annotations in domains with scarce labeled data.