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
Results
Objective Evaluation of Network Performance:
1005 control and 193 case slice images met inclusion/exclusion criteria (Fig. 1).
After 5 runs of 200 epochs of training,
training accuracy averaged 0.893 with average validation accuracy equaling 0.853.
Multiple cross-fold training and validation runs produced validation accuracy values ranging from 0.828 - 0.877.
Average training time per epoch was 36 seconds on a single Titan X Pascal (Nvidia,
Santa Clara,
California).
Average time to classification inference per CT head volume was under 3 seconds.
Average time to the generation of the heat map visualizations (i.e.
Salience Map,
Grad-CAM) averaged 28 seconds on initial processing per image slice.
Subjective Evaluation of Network Visualization Maps:
Upon review of the generated Salience and Grad-CAM visualization patterns we noted differences in the level of localization precision with Salience Maps favouring high-frequency pinpoint features more suitable for SAH localization versus larger non-specific regions of highlight for Grad-CAM (Fig. 2).
Grad-CAM visualizations would often propose regions too large to specifically identify foci of bleed and oftentimes the regions proposed would fail to even remain well-centered on reasonable proposals (Fig. 3).
Common areas of false positive highlight include a hyperdense posterior falx,
as well calvarial inner table calcifications or activation of the calvarium itself (Fig. 4).