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
Conclusion
Future Directions
Training and validation of the networks were limited by the low quantity of cases and controls,
and by the severe class imbalance in the case of pneumoperitoneum.
Given the relative lack of anatomical conservation for small-bowel obstruction,
a larger dataset or improved engineering methods for dealing with small and imbalanced datasets are needed.
We propose multiple avenues of engineering development within the context of computer-assisted triaging for acute conditions.
In order to resolve the issue of data scarcity,
especially for uncommon or rare conditions,
we believe that using generative adversarial networks (GANs),
as opposed to traditional affine data augmentation,
can yield improved detection and localization performance.
Though specific implementations vary widely,
GANs are dual or multi-network adversarial architectures that attempt to generate artificial data distributionally indistinguishable from its training cases [12],
and can thus be used to generate complex data for use cases such as data augmentation [13].
With respect to network performance,
adopting object detection networks such as RetinaNet could allow for explicit localization of acute conditions on thoracic and abdominal radiographs at higher resolutions [14].
As shown by the lack of resolution on small-bowel obstruction detection,
existing methods of using saliency mapping and Grad-CAM especially suffer in terms of finding fine-grained features on radiographs.
This approach comes however at a cost,
as object detection networks like RetinaNet require the manual input of rectangular regions of interest (ROI) or segmentation contours,
thus further limiting the data available for training.
Localization networks are particularly impractical when viewed within the context of a diagnostic workflow for acute conditions.
The systematic implementation of such an approach would depend upon the integration of active learning tools within existing PACS systems that could incorporate localization and segmentation metadata alongside studies.
Conclusion
We propose an automated pipeline for triaging of acute thoracic and abdominal events on radiographs,
focusing on pneumoperitoneum and small-bowel obstruction respectively.
By using integrations in automated PACS extraction using keyword searching,
negation detection,
DICOM modality and view extraction,
multiple condition-specific neural networks are able to accurately detect acute findings and adequately localize them to relevant parts of the body.
Further developments along these lines may realize over time the potential of automated triaging of acute findings on thoracic and abdominal X-rays.