Keywords:
Artificial Intelligence, CNS, CT, Computer Applications-Detection, diagnosis, Computer Applications-General, Trauma, Aneurysms
Authors:
W. Ding, J. W. Luo, J. J. R. Chong; Montreal, QC/CA
DOI:
10.26044/ecr2019/C-3184
Conclusion
Study Limitations:
Both training and validation were limited by the relative low quantity of positive training images as well as the large number of discarded training cases due to multi-compartmental bleed.
In general,
as with previous studies,
we found the relative value of reported ROC accuracy figures to be overly optimistic with respect to accuracy with variable performance when attempting subjective visual validation.
Because of the low dataset size,
alternate split designs and validation strategies such as cross-validation had to be employed to obtain a more realistic estimate of network performance and could benefit from either further expansion of the dataset,
or multi-institutional datasets.
Future Directions:
The results from these experiments suggest the need for multiple avenues of engineering development.
From the results of this preliminary experiment,
we feel there is particular room for improvement in employing more advanced network designs that permit training with more precise localization to help exclude normal regions of head anatomy which should reduce the number of false positives on localization visualization.
In particular,
we also feel that existing ImageNet-style visualizations that have worked well for well-centered or low-frequency features such as Grad-CAM (i.e.
objects that occupy >20% field-of-view) may be extremely inappropriate for fine-grained findings like subarachnoid hemorrhage.
While class-specific heat maps are often a useful function in traditional computer vision applications,
the disadvantages with respect to spatial resolution create significant clinical limitations.
Finally,
given the nature of the false positive activations found on visualization,
the application of more classical computer vision region-based or structural mapping/segmentation techniques may aid in reducing the number of false-positive activations.
Conclusion:
Convolutional neural networks utilizing ResNet architectures can successfully classify and localize bleeds with moderate accuracy.
High-frequency visualization methods such as Salience Maps are better suited for the relatively smaller finding size of SAH and permit more precise finding localization.