Keywords:
CAD, Conventional radiography, Lung, Artificial Intelligence, Image verification
Authors:
M. Grass, I. M. Baltruschat, A. Saalbach, H. Nickisch, J. von Berg, H. Ittrich, L. Steinmeister, G. B. Adam, T. Knopp; Hamburg/DE
DOI:
10.26044/ecr2019/C-1093
Methods and materials
In this study,
we employ bone suppression [3],
an algorithm to artificially remove the rib cage in chest X-ray images,
and automatic lung field detection to crop images to the lung area,
in order to study their usefulness in the context of Deep Learning.
Furthermore,
we consider the combination of both.
For Convolutional Neural Network (CNN) training and evaluation,
DICOM images from the Indiana dataset (Open-I)[1] were examined by two expert radiologists and annotated with respect to eight different pathologies: pleural effusion,
infiltrate,
congestion,
atelectasis,
pneumothorax,
cardiomegaly,
mass,
foreign object.
Following the method and training setup in [4],
we pre-trained a dedicated ResNet-50 architecture with a larger input size of 448×448 (see Fig.
2) on ChestX-ray14,
the largest publicly available X-ray dataset and fine-tuned it by using the Open-I data. For network evaluation,
we resized images to 480×480 and employed an average five crop for evaluation.