Skip to main content
Insights into Imaging
Fig. 5 | Insights into Imaging

Fig. 5

From: Weakly supervised segmentation models as explainable radiological classifiers for lung tumour detection on CT images

Fig. 5

Model explainability heatmaps. The first five positive test instances are shown. Models were trained to detect NSCLC tumours at image level. WSUnet’s heatmap was extracted from the penultimate voxel-level output layer. GradCAM heatmaps were extracted from the seventh (“GradCAM (16, 16, 64)”) and ninth (“GradCAM (8, 8, 128)”) convolutional layers. Nearest-neighbour interpolation was applied to map GradCAM, integrated gradients and occlusion sensitivity heatmap to the input image dimensions. For comparability, methods heatmaps were normalised to the range of minimum and maximum values for the five images

Back to article page