Using Cooperative Multi-Agent Reinforcement Learning for Mammogram ROI Classification
Md. Sazid Uddin, M. F. Mridha, Md. Abdullah-Al-Jubair · 2023
This paper explores the use of cooperative Multi-Agent Reinforcement Learning (MARL) for identifying cancerous cells in annotated regions of interest in mammogram images with the goal of early detection and treatment which is key to reducing the risk of cancer development and death. Our proposed MARL model deploys multiple agents on mammogram images that collaboratively scan and communicate observations with other agents to extract features in a decentralized fashion and classify ROI patches as benign or malignant. As opposed to processing whole images at a time, partial observations by agents result in decreased computational complexity. The achieved accuracy, precision, and recall of our architecture on the CBIS-DDSM (mass) dataset were 82.45%, 81.73% and 81.00% respectively.