Fig. 3: Prognostic prediction performance. | Communications Medicine

Fig. 3: Prognostic prediction performance.

From: A convolutional attention model for predicting response to chemo-immunotherapy from ultrasound elastography in mouse tumor models

Fig. 3

a Macro-average ROC with the one-vs-rest scheme that shows the performance metrics difference between the Prognose-CNNattention model and the pre-trained networks (right) and the performance metrics difference between the Prognose-CNN models with and without attention, and Auto-Prognose-CNNattention framework (SegforClass) (left). b ROC curves of the three classes for the Prognose-CNNattention model. The performance of a laboratory experts is plotted in a, b by twins cross/star respectively across two sessions. c Confusion matrix of the Prognose-CNNattention. d Prognose-CNNattention model performance per treatment group. The ratio of responsive, stable and non-responsive tumors per treatment group in our dataset compared to how the model classified the tumors in each treatment group. P values for the ROC curves between the ROC of the Prognose-CNAttention and the ROC of best pre-trained (Prognose-CNN models without attention, Auto-Prognose-CNNattention, Xception, VGG16, Inception-V3, ResNet500.043, 0.0021, 0.0015, 0.0001, 0.0007. 0.0023 respectively.

Back to article page