Zero-Shot Medical Image Retrieval for Emerging Infectious Diseases Based on Meta-Transfer Learning - Worldwide, 2020

Introduction: Due to the increasing number of medical images, image retrieval has become an important technique for medical image analytics. Although many content-based image retrieval methods have been proposed, the retrieval of images in datasets related to emerging/new infectious diseases still remain a challenge-mostly due to the lack of historical data. As a result, the current retrieval models have limited functionality in helping doctors make accurate diagnoses of new diseases. Methods: In this paper, we propose a zero-shot retrieval model based on meta-learning and ensemble learning, which can obtain a model with stronger generalizability without using any relevant training data, and thus performs well on new types of test data. Results: The experimental results showed that the proposed method is 3% to 5% higher than the traditional method, which means that our model can retrieve relevant medical images more accurately for newly emerging data types and provide doctors with more effective assistance. Discussion: When a new infectious disease occurs, doctors can use the proposed zero-shot retrieval model to retrieve all relevant cases, quickly find the common problems of patients, find the locations of the new infections, and determine its infectivity as soon as possible. The proposed method is a new computeraided decision support technology for emerging infectious diseases.