Fig. 2From: Weakly supervised segmentation models as explainable radiological classifiers for lung tumour detection on CT imagesModel training and evaluation schema. Image patches were sampled in the MAASTRO dataset, and labelled at image level according to whether the image contained any tumour-positive voxels. Model training utilised the image-level labels only; the model was not provided with any information on tumour location. The model evaluation used both image-level labels and voxel-level ground truthBack to article page