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
Aims and objectives
Acute chest and abdominal conditions,
namely pneumoperitoneum and small-bowel obstruction,
are urgent and critical findings found on chest and abdominal radiographs that frequently require prompt medical or surgical intervention.
Small bowel obstruction alone accounts for 12 to 20% of surgical admissions for acute abdominal pain and is a significant cause of morbidity and mortality [1][2].
Yet,
timely interpretation for X-rays remains a significant challenge due to the high volume of cases,
low prioritization compared to MRI,
CT and ultrasound,
and limited time and human resources available.
From Medicare data,
Rao et al.
report a 227% increase in CT utilization during the 8 year period from 2000 to 2008,
while radiograph utilization only increased by 29% [3].
Due to the urgency of the aforementioned conditions and correspondingly low available resources to interpret them,
there is a strong need for a triaging system that could intelligently prioritize X-rays and perform preliminary reading to screen for acute events.
The integration of a computer-aided diagnosis (CAD) and triaging pipeline within the diagnostic workflow could improve radiologist efficiency,
reduce the time to interpretation for acute conditions,
and thus lead to improved patient outcomes in terms of mortality and morbidity.
A number of groups have recently demonstrated the ability for convolutional neural network-based (CNN) computer-assisted diagnosis to detect acute chest and abdominal conditions.
Using an Inception-v3 network,
Cheng et al.
obtained an area under the curve (AUC) of 0.84 for the detection of small bowel obstruction,
analyzed on 3663 supine abdominal radiographs [4].
A further refinement of that model by Cheng et al.
obtained an AUC of 0.803 with 2210 radiographs,
and of 0.971 with 7768 radiographs [5].
Though no previous studies pertaining specifically to the automated detection of pneumoperitoneum has been found,
research on the Chest-Xray14 dataset released by Wang et al.
has demonstrated the potential of deep learning for the detection of chest conditions [6].
AUC values of the provided ResNet50 ranged from 0.60 to 0.83 on 14 conditions,
with two of them being acute (pneumothorax and hernia).
Improving on that initial model,
the DenseNet-based CheXNet model by Rajpurkkar et al.
obtained AUC values ranging from 0.73 to 0.92 [7].
As shown by existing model performance,
deep learning-based computer-aided diagnosis has strong and growing potential towards the automated detection of acute thoracic and abdominal conditions.
As such,
we propose a deep learning-based triaging pipeline on thoracic and abdominal radiographs in order to automatically triage and perform preliminary diagnosis for two acute findings,
pneumoperitoneum and bowel obstruction.
Inspired by the end-to-end triaging system for acute neurological events proposed by Titano et al.
[8],
we extend existing approaches for computer-aided diagnosis in radiography traditionally focused on imaging alone by incorporating automated keyphrase-based extraction and view detection components while using latest deep convolutional neural network (CNN) architectures for the classification component.