Skip to main content
. 2020 Nov 17;8(11):e19805. doi: 10.2196/19805

Table 3.

Performance of the deep neural network (DNN) models on individual magnetic resonance imaging (MRI) sequences: T1-weighted MRI (T1), T2-weighted MRI (T2), and gadolinium-contrast-enhanced T1-weighted MRI (T1c).

DNN model and magnetic resonance sequence Sensitivity (95% CI) Specificity (95% CI) Accuracy (95% CI) Area under the curve (95% CI)
VGGa16




T1 0.725 (0.696-0.753) 0.606 (0.562-0.648) 0.684 (0.660-0.708) 0.718 (0.689-0.747)

T2 0.690 (0.660-0.719) 0.686 (0.644-0.727) 0.689 (0.665-0.713) 0.767 (0.740-0.794)

T1c 0.874 (0.851-0.894) 0.540 (0.496-0.585) 0.759 (0.736-0.781) 0.770 (0.743-0.797)
VGG19




T1 0.804 (0.778-0.829) 0.448 (0.404-0.492) 0.681 (0.657-0.705) 0.692 (0.663-0.721)

T2 0.743 (0.714-0.770) 0.554 (0.510-0.598) 0.678 (0.653-0.702) 0.741 (0.713-0.769)

T1c 0.800 (0.773-0.825) 0.653 (0.610-0.694) 0.749 (0.726-0.771) 0.795 (0.769-0.821)
ResNetb-50




T1 0.782 (0.755-0.808) 0.584 (0.540-0.627) 0.714 (0.690-0.737) 0.732 (0.704-0.760)

T2 0.833 (0.808-0.852) 0.525 (0.480-0.569) 0.727 (0.703-0.750) 0.762 (0.735-0.789)

T1c 0.825 (0.799-0.848) 0.653 (0.610-0.694) 0.766 (0.743-0.787) 0.824 (0.800-0.848)
Inception-v3




T1 0.724 (0.695-0.752) 0.596 (0.552-0.639) 0.680 (0.656-0.704) 0.706 (0.677-0.735)

T2 0.634 (0.603-0.665) 0.734 (0.693-0.772) 0.668 (0.644-0.693) 0.734 (0.706-0.762)

T1c 0.769 (0.741-0.795) 0.732 (0.691-0.770) 0.756 (0.733-0.778) 0.831 (0.807-0.855)
Inception-ResNet-v2




T1 0.774 (0.746-0.800) 0.590 (0.546-0.633) 0.711 (0.687-0.734) 0.748 (0.720-0.776)

T2 0.829 (0.804-0.852) 0.529 (0.484-0.573) 0.726 (0.702-0.748) 0.804 (0.779-0.829)

T1c 0.812 (0.786-0.837) 0.722 (0.681-0.761) 0.781 (0.759-0.802) 0.841 (0.818-0.864)
ERN-Netc




T1 0.704 (0.674-0.732) 0.519 (0.474-0.563) 0.640 (0.615-0.665) 0.646 (0.615-0.676)

T2 0.634 (0.603-0.665) 0.606 (0.562-0.648) 0.624 (0.599-0.649) 0.675 (0.645-0.705)

T1c 0.803 (0.777-0.828) 0.643 (0.600-0.685) 0.748 (0.725-0.770) 0.807 (0.782-0.832)

aVGG: Visual Geometry Group.

bResNet: residual neural network.

cERN-Net: efficient radionecrosis neural network.