Keywords:
Artificial Intelligence, Lung, Anatomy, Digital radiography, Diagnostic procedure, Infection
Authors:
L. Rusko1, A. Radics2, V. Venugopal3, K. Nye4, G. Avinash3, V. Mahajan3; 1Szeged/HU, 2Budapest/HU, 3New Delhi/IN, 4Milwaukee, WI/US
DOI:
10.26044/ecr2019/C-2327
Methods and materials
Image dataset:
The dataset involved 3531 XRAY images (1024x1024),
which were randomly separated into (2714) training,
(342) validation,
and (475) test sets,
such that each patient had images in one set only.
The chest-cavity and lung masks were available for all cases.
Model training: The XRAY images were resampled to 255x255 resolution and min-max normalization was applied to the (8-bit,
grayscale) intensity.
The model architecture is UNET.
(The contracting part of the network involves concatenated blocks of convolution,
batch-normalization and dropout,
the result of that is down-sampled by a factor of 3.
The expanding part involves concatenated blocks of transpose convolution,
batch-normalization and dropout,
the result of that is up-sampled by a factor of 3.) The model has 120 layers and 2,413,762 trainable parameters.
The model was trained for 500 epochs (with batch size equal to 10).
After all augmented (random shift,
rotation and scaling) training samples were processed in an epoch,
the model was evaluated on the validation samples and its parameters were saved if the accuracy (DICE score) was improved.
Separate models were trained for chest-cavity,
left- and right lung.
Model testing:
The best model (of 500 epochs) was evaluated on the test samples which were down-sampled (to 255x255 resolution) and normalized (min-max) before the model was applied to them.
The output of the model was rounded (to binary image) and compared with the reference mask (that was down-sampled to model resolution) using DICE score (recall and precision).