Knowledge Extraction and Discrimination Based Calibration on Medical Imaging Classification

The calibration of modern deep learning methods is often neglected when they are applied to medical diagnosis. Meanwhile, the effectiveness of traditional calibration methods heavily relies on the size of validation set, which is not suitable for scenarios with limited medical images. To this end, we propose a knowledge extraction and discrimination model (KED), in which the discrimination network is set to generate fine-grained pseudo-labels based on the information of the training process; the classification network is trained with both binary classification and fine-grained pseudo-labels. The calibration of the model is achieved during the training process by explicit modeling. Experiments are conducted on breast MRI images based on BI-RADS assessment categories, and the results show that our model achieves the best overall calibration level and the classification accuracy is also improved compared with other calibration methods.