Multi-label neural architecture search for chest radiography image classification

Yi Yang, Jiaxuan Wei, Zhixuan Yu, Ruisheng Zhang

Abstract: Chest radiography remain the global standard for diagnosing pulmonary diseases. Despite numerous research efforts, medical professionals still face challenges in rapidly and accurately analyzing multiple diseases on a single chest radiography. Moreover, traditional deep learning methods suffer from complexities in design and prolonged processing times. To address these issues, we propose a multi-label neural architecture search (MLNAS) approach. Primarily intended for multi-label chest radiography image classification, MLNAS employs automated modeling, data augmentation, and threshold calculation strategies to improve the accuracy of chest radiography image classification and enhance result interpretation. Furthermore, MLNAS demonstrates the potential for application in other multi-label medical image classification domains. Experimental results indicate that MLNAS achieves state-of-the-art prediction accuracy for 9 out of 14 lung diseases. This novel approach presents a new solution for computer-aided diagnosis of chest X-rays.