Figure 3 | Scientific Reports

Figure 3

From: A joint convolutional-recurrent neural network with an attention mechanism for detecting intracranial hemorrhage on noncontrast head CT

Figure 3

A 68-year-old female with known hypertension (The images were created by the authors using open-source software, Matplotlib v3.5, Python v3). A right thalamic hematoma extended into the adjacent ventricular system on a non-contrast head CT scan (right). NormGrad (middle) method generates more delicate saliency maps than Grad-CAM (left), highlighting the thalamic hematoma and its ventricular extension. The average quality scores were 3.6 points and 2 points for the NormGrad and Grad-CAM, respectively. Please note that the observers evaluated saliency maps with the same color spectrum, and the current color maps are adjusted for representative purposes.

Back to article page