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

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Supervised learning models in computer vision require large labeled datasets, which are expensive and time-consuming to create.
• Automating the annotation process is crucial for efficient training data preparation.

Purpose of the Study:

• To present a Deep Reinforcement Learning (DRL) based approach for automatic annotation generation, reducing human effort in supervised learning.
• To introduce the "Try a Little More" (TLM) methodology, inspired by constructivist teaching, to enhance agent learning through human guidance and active learning.

Main Methods:

• Developed a virtual agent trained with human guidance using constructivist teaching principles.
• Implemented active learning within the TLM approach to identify uncertain cases and request human intervention.
• Evaluated the agent's ability to autonomously create bounding box annotations on a mammography dataset.

Main Results:

• The TLM approach generated 414 new annotations with high IoU (0.86) and F1-score (0.92).
• Utilizing TLM annotations significantly improved a YOLO detector's mAP from 0.52 to 0.91 (75% improvement).
• Achieved a 35% reduction in human intervention for annotation generation.

Conclusions:

• The proposed TLM methodology effectively accelerates annotation creation and enhances training data quality.
• This approach advances Data-Centric AI by combining human advice with reinforcement learning for efficient data annotation, especially in data-scarce domains.