Keywords:
Artificial Intelligence, Digital radiography, Neural networks, Computer Applications-Detection, diagnosis, Acute
Authors:
J. W. Luo, J. J. R. Chong; Montreal, QC/CA
DOI:
10.26044/ecr2019/C-3491
Methods and materials
Data Selection:
Automated RIS keyphrase searches specific for acute chest and abdominal conditions were performed on all radiological reports from 2006-2017 on PACS.
Alongside negation detection,
the keyphrase extraction automatically identified 84,399 potential studies (Fig. 1).
A DICOM-parsing algorithm was then applied to filter each study for non-radiograph modalities,
and then used to classify radiographs under posteroanterior (PA),
anteroposterior (AP),
lateral view for chest X-rays,
and supine,
upright or decubitus views for abdominal X-rays.
Refining the search to only include reports suggestive for pneumoperitoneum and small-bowel obstruction,
14,336 studies were kept.
This corresponds to 968 X-rays positive for pneumoperitoneum alongside 9,783 controls in PA and AP view,
and 1,565 X-rays suggestive for abdominal obstruction alongside 2,020 controls in upright and supine view (Fig. 1).
Lateral or decubitus views seen in studies were discarded as their inclusion did not seem to improve model performance.
Network Configuration:
Each thoracic and abdominal detector was separately trained using the same Inception-ResNet-v2 convolutional neural network (CNN) architecture [9].
Each dataset was split using standard 70% training,
10% validation,
and 20% splits.
Networks were trained using stochastic gradient descent (SGD) with an initial learning rate of 0.003 alongside weight decay and 5 x 20-epoch cosine annealing schedules for a total of 100 epochs [10].
We used a weighted binary cross-entropy loss function with per-class scaling in order to address class imbalance,
resulting in cases being weighted 10.1x and 1.29x more than controls for pneumoperitoneum and small-bowel obstruction respectively.
Implementation of the neural network was done under TensorFlow on Python 3.6,
and training was performed on a Titan X (Pascal) workstation.
We performed transfer learning from pre-trained weights from the 2012 Large Scale Visual Recognition Challenge dataset (ImageNet) alongside data augmentation consisting of standard affine transformations (random cropping,
rotation,
shearing,
and horizontal flipping).
All radiographs underwent histogram normalization prior to training.
Performance evaluation:
Quantitative network performance was assessed based on classification accuracy and area under the receiver operating characteristic (AUC) curves.
A further qualitative review of the generated saliency maps and gradient-weighted class activation maps (Grad-CAM) was performed in order to review the localization performance of the networks [11].