Convolutional neural network to detect thorax diseases from multi-view chest X-rays

Abstract

Chest radiography is the most common examination for a radiologist. This demands correct and immediate diagnosis of a patient’s thorax to avoid life threatening diseases. Not only certified radiologists are hard to find, stress, fatigue and experience contribute to the quality of an examination. It is ideal that a chest X-ray can be interpreted by an automated deep learning algorithm. In this paper, we proposed a stage-wise model that is founded on a ResNet-50 based deep convolutional neural networks architecture to detect the presence and absence of twelve thorax diseases. This novel model has incorporated various recent techniques such as transfer learning, fine tuning, fit one cycle function and discriminative learning rates. The experiments were performed on 10% of the largest collection of chest X-rays to date, the MIMIC-CXR dataset. The model was trained for eight epochs using a subset of the available multi-view chest X-rays. The absolute labelling performance has achieved an encouraging average AUC of 0.779.