Domain adaptation from posteroanterior to anteroposterior X-ray radiograph classification via deep neural converter with label recycling

Deep neural network trained with massive labeled dataset has shown impressive performance in multiple disease classification tasks for chest X-ray radiograph (CXR). However, different imaging protocols even for the same disease such as posteroanterior (PA) and anteroposterior (AP) CXR often lead to domain gap, causing significant performance degradation and requiring more images with painstaking labeling. We propose a novel domain adaptation scheme via deep neural converter (DNC) in the scenario of having labeled PA and unlabeled AP; converting labeled PA into AP with label recycling and pseudo-labeling unlabeled AP. To overcome no labeled AP data, our proposed method exploits DNC to convert labeled PA CXR into AP CXR with label recycling as well as pseudo labeler for unlabeled AP CXR.

I. Introduction

Recently, deep neural networks trained with massive labeled dataset has shown impressive performance in computer vision such as classification [12], [16] and multiple disease classification [1]–[4]. However, deep neural networks cause significant performance degradation when there is a gap between the domains of the training data and the test data. In order to reduce the domain gap, many studies have been conducted in the field of domain adaptation in computer vision [5]–[8].