基于高维多目标优化的小样本皮肤癌检测

Few-shot skin cancer detection based on many-objective optimization

由于皮肤癌数据的长尾分布特性,快速识别含少量数据的罕见皮肤病样本成为一个具有挑战性的小样本问题.基于元学习的检测方法能够从多数常见皮肤病样本中快速学习元知识,利用先验知识提高模型检测罕见皮肤病的能力.然而,皮肤癌公共类别质量和分布的偏差,导致元学习在预训练阶段存在过拟合风险,且基于传统网络的元学习模型难以处理细粒度皮肤病问题.针对此问题,提出一种高维多目标元学习皮肤癌检测模型.该模型在元学习的基础上,通过考虑皮肤癌检测模型的多种分类性能,优化公共类别(基类)分布获得强化的训练样本;采用融合CCNet注意力机制的ResNet12网络结构,充分提高识别细粒度皮肤病变图像的能力.此外,设计一种基于离散分组交叉策略的高维多目标优化算法对所提出的高维多目标皮肤癌检测模型进行高效求解.在ISCI2018和Derm7pt两个公开的医学数据集上进行实验,在二分类的1次、3次和5次采样任务中,分别获得67%、79%、82%的检测准确率,验证了高维多目标检测皮肤癌检测模型的有效性.

Due to the long-tailed distribution characteristics of skin cancer data,quickly identifying rare skin disease samples with small amounts of data becomes a challenging few-shot problem.The detection method based on meta-learning can quickly learn meta-knowledge from the majority of common classes skin disease classes and use prior knowledge to improve the model’s ability to detect rare skin diseases.However,the bias in the quality and distribution of public skin cancer categories leads to the risk of overfitting in the pre-training phase of meta-learning,and meta-learning models based on traditional networks struggle to handle fine-grained skin disease problems.Affronting this issue,a many-objective meta-learning skin cancer detection model is proposed.The proposed model optimizes the distribution of common classes(base classes)by considering various classification performances of the skin cancer detection model,thus obtaining quality-enhanced training samples.And it adopts the ResNet12 network structure integrated with the CCNet attention model,which can significantly enhance the ability to identify fine-grained skin lesion images.Additionally,an improved many-objective optimization algorithm with discrete grouping strategy is designed to efficiently solve the proposed model.Some experiments are conducted on two public medical datasets,ISCI2018 and Derm7pt.The proposed many-objective skin detection model achieves detection accuracies of 67%,79%,and 82%respectively on binary classification tasks with 1-shot,3-shot,and 5-shot learning,which validates the effectiveness of the model.