Article Figures & Data
Figures
- FIG 1.
Head CT scan slices annotated as ICH (upper and middle, arrows) and annotated as no-ICH (bottom).
- FIG 2.
Flowchart of the study. The upper left corner shows the base architecture underlying all 70 CNNs, which was then structurally optimized by the 70 teams, resulting in 70 different CNNs. These were trained on data sets 1 and 2 and were used to train the 4 ensemble learning methods (MV indicates majority voting; DT, decision tree; MI-DNN, multi-input deep neural network; see text) on data set 3. Both the 70 trained CNNs and the trained ensemble methods were tested on data set 4. ReLU indicates rectified linear unit; tanh, hyperbolic tangent; EL, ensemble learning.
- FIG 3.
Accuracy of individual CNNs and ensemble learning methods. The purple solid vertical lines show the accuracy achieved by each ensemble learning method, and the black dashed vertical lines show means and lower and upper 95% confidence intervals for the distribution of accuracy of individual CNNs. Additionally, the first purple solid line indicates the accuracy of the base network. DNN indicates deep neural network.
- FIG 4.
Receiver operating characteristic curves of the base network (dark blue) and the single best CNN (dark red), the curve of the average performing CNN (orange), and ensemble learning methods. The dashed black line indicates the performance of a random classifier (ie, accuracy of 50%). DNN indicates deep neural network.
Jump to section
Related Articles
Cited By...
- No citing articles found.