TriageNet: A Multi-Agent Diagnosis Network for Imbalanced Data

Abstract

Imbalanced and even long-tail distribution of different categories is a challenge for multi-class classification problem, especially for medical image diagnose whose data distribution is usually imbalanced. Toward this issue, we proposed an end-to-end multi-agent classification network called Tria-geNet, which is combined of multiple selectors and diagnostic agents. All categories are guided to different agents by selectors, and every agent is an expert in a specific group of categories. This process, which is similar to triage in hospitals, helps decrease the unbalance between categories for both selectors and agents. Experiments on an extremely imbalanced pneumonia CT dataset and a publicly available X-ray dataset Chexpert show that TriageNet is relatively robust to imbalanced data.