FIGURE 8.

The flow of the entire method. Among them, we slice the entire pathological data and extract the effective diseased area as much as possible. The active learning strategy follows our work in Qi et al. (2018). The goal of that work is to maximize learning accuracy from very limited labeling data. The classification model is updated iteratively with an increasing training set. The sliced pathological data are sent to a convolutional neural network to obtain the discrimination results of the pathological data. The radiological data are sent to the Unet and CNN to obtain classification results after preprocessing. Finally, the results are combined via a weighted average operation to obtain the final result.