Keywords:
Infection, Computer Applications-General, Digital radiography, Conventional radiography, Respiratory system, Artificial Intelligence
Authors:
T. Fung1, J. W. Luo2, T. C. Lee3, B. Gallix2, J. J. R. Chong2; 1Montreal, Québec/CA, 2Montreal, QC/CA, 3Montreal/CA
DOI:
10.26044/ecr2019/C-3416
Methods and materials
Using a keyphrase search approach,
a set of RIS search queries were created to identify chest radiograph studies highly suspicious for tuberculosis.
Frontal view DICOM images were exported.
Using this search,
205 frontal images of radiographically suspicious tuberculosis were identified,
with 301 of 820 quadrants containing disease.
8,933 normal chest radiographs were used as negative control.
We employ a dual-channel ImageNet pre-trained deep convolutional neural network (DCNN) derived from the Inception-v2-Resnet architecture to detect tuberculosis (Figure 1).
The dataset was split using 70% training,
10% validation,
and 20% testing ratios.
Using stochastic gradient descent (SGD) and an initial learning rate of 0.003 alongside weight decay,
we trained the network for a total of 100 epochs.
To address the inherent class imbalance,
we employ per-class scaling on the binary cross-entropy loss function,
which results in a 10.11x higher penalty if the network made an incorrect classification.
The trained networks were evaluated using areas under the receiving operating characteristic curves using DeLong’s nonparametric method.