Figure 2

The identification of bladder tumor stage using the fine-tuned CNNs and urologist ratings. (A) The class sensitivities and specificities of the considered CNNs and both physician ratings as a two-directions bar chart. While both urologists could not assess well the last two invasive stages of bladder cancer, the detection of these tumor stages was quite good based on all CNNs (B) The obtained mean sensitivities and mean specificities for all classification models. The best classification results were achieved by the MobileNetV2 network with a mean sensitivity of 88% followed by the mean sensitivity obtained by the InceptionV3 network. The classification mean sensitivity decreased at least 50% when the same images were assessed by the urologists. (C) The class distribution of the BL image data set. The number of involved images for this task varies a lot from one class to another class.