Application of Convolutional Neural-Network with Uncertainty to Pneumoconiosis Classification

Abstract

Uncertainty-based recognition system is proposed to perform pneumoconiosis image detection using convolutional neural networks (CNNs). The Bayesian network based on ResNet18 is used as the CNN, which enables to create a CNN considering uncertainty. As a further development, we improved the accuracy by using ensemble learning. We evaluated three types of models: the accuracy of the four created models one-by-one, ensemble learning considering the uncertainty of the models, and ensemble learning considering the uncertainty of the recognition results. As a result, the highest accuracy in the model alone is 82.49%. The softmax value, which is a measure of uncertainty in the image, and the standard deviation of the softmax value after 100 predictions with dropout show different values for the softmax value, and the standard deviation was found to be close to zero for all models. In the case of ensemble learning with the uncertainty of the models, the accuracy was 88.88% when the larger value of the softmax values in the label of the final output layer of the model is added. In the case of ensemble learning with the uncertainty of the recognition result, the accuracy was 86.11% when the standard deviations of the softmax values were added together after 100 iterations of prediction and the larger value was used as the final result. This result suggests that the accuracy of ensemble learning is higher than that of the stand-alone method.