基于预测对抗网络的图像二分类模型

正未标记学习仅使用无标签样本和正样本训练一个二分类器, 而生成式对抗网络(generative adversarial networks, GAN)中通过对抗性训练得到一个图像生成器. 为将GAN的对抗训练方法迁移到正未标记学习中以提升正未标记学习的效果, 可将GAN中的生成器替换为分类器C, 在无标签数据集中挑选样本以欺骗判别器D, 对C与D进行迭代优化. 本文提出基于以Jensen-Shannon散度(JS散度)为目标函数的JS-PAN模型. 最后, 结合数据分布特点及现状需求, 说明了PAN模型在医疗诊断图像二分类应用的合理性及高性能. 在MNIST, CIFAR-10数据集上的实验结果显示: KL-PAN模型与同类正未标记学习模型对比有更高的精确度(ACC)及F1-score; 对称化改进后, JS-PAN模型在两个指标上均有所提升, 因此JS-PAN模型的提出更具有合理性. 在Med-MNIST的3个子图像数据集上的实验显示: KL-PAN模型与4个benchmark有监督模型有几乎相同的ACC, JS-PAN也有更高表现. 因此, 综合PAN模型的出色分类效果及医疗诊断数据的分布特征, PAN作为半监督学习方法可获得更快、更好的效果, 在医学图像的二分类的任务上具有更高的性能.

Medical Image Classification Based on Predictive Adversarial Networks

Positive-unlabeled learning (PU learning) only uses unlabeled samples and positive samples to train a binary classifier, while generative adversarial networks (GANs) obtain an image generator through adversarial training. In order to transfer the adversarial training method of GANs to PU learning for higher PU learning performance, the generator in GANs can be replaced with a classifier C, which selects samples in the unlabeled dataset to deceive the discriminator D and optimize C and D iteratively. This study proposes the JS-PAN model, which uses the Jensen-Shannon divergence (JS-divergence) as the objective function. Finally, according to the characteristics of data distribution and current needs, the rationality and high performance of the PAN model applied in the binary classification of medical diagnostic images are explained. Experiments on MNIST and CIFAR-10 datasets show that the KL-PAN model has higher accuracy (ACC) and F1-score than the similar PU learning models, and the JS-PAN model has higher performance in terms of two indicators after symmetric improvement, so the JS-PAN model is more reasonable. Experiments on three image subdatasets of Med-MNIST show that the KL-PAN model has almost the same ACC as the four benchmark supervised models, and JS-PAN has higher performance. Therefore, in view of both the excellent classification performance of the PAN model and the distribution characteristics of medical diagnostic data, PAN, as a semi-supervised learning method, can achieve faster and better results and thus show higher performance in the task of binary classification of medical images.