基于深度贝叶斯蒸馏网络的皮肤病变识别研究

Research on skin lesion recognition based on deep Bayesian distillation network

目的:解决卷积神经网络无法量化模型的偶然不确定性与认知不确定性问题,优化皮肤病变类型识别机制,降低误诊率。方法:在构建多尺度网络时,基于贝叶斯深度学习网络构建深度贝叶斯蒸馏网络,通过多次采样数据分布方式拟合训练数据模型,量化模型的偶然不确定性与认知不确定性。进一步引入了知识蒸馏对模型进行压缩,构建学生网络模型拟合教师网络的输出,使用教师网络的参数和真实值标签训练学生网络,从而实现对模型参数量与时间的优化。结果:识别准确率与现有相关方案相比提高3.00%~8.00%,达到83.90%,同时参数量减少14.12%,运行时间节约8.70%。结论:基于深度贝叶斯蒸馏网络的皮肤病变识别机制能够显著提高识别准确率,同时减少模型参数量与运行时间。

Objective In order to solve the problem of stochastic and cognitive uncertainty in models that cannot be quantified by convolutional neural networks,optimize the recognition mechanism of skin lesion types and reduce the misdiagnosis rate.Methods When constructing the multi-scale network,the deep Bayesian distillation network was constructed based on the Bayesian deep learning network,and the stochastic and cognitive uncertainty of the model were quantified by fitting the training data model with multiple-sampling of data distribution.Furthermore,knowledge distillation was introduced to compress the model,a student network model was constructed to fit the output of the teacher network,and the parameters and true value labels of the teacher network were used to train the student network,so as to optimize the number of model parameters and running time.Results Compared with the existing schemes,the recognition accuracy was increased by 3.00%~8.00%to 83.90%,while the number of parameters was reduced by 14.12%,and the running time was saved by 8.70%.Conclusion The recognition mechanism of skin lesions based on deep Bayesian distillation network can significantly improve the recognition accuracy,and reduce the number of model parameters and running time.