基于对比学习多级特征融合的尿路上皮细胞分类方法
Urinary Epithelial Cell Classification Method Based on Comparative Learning and Multilevel Feature Fusion
-
摘要: 为应对当前细胞分类方法易受图像质量影响以及标注数据稀缺而导致分类准确率不高的问题, 提出一种基于多级特征融合的混合自监督学习方法. 首先, 多级特征融合模块可在特征提取时同时提取语义特征信息和空间特征信息, 以利于识别不同特征形态的阳性细胞; 其次提出自适应对比损失函数将自监督学习应用在尿脱落细胞分类中, 以促进模型可最大化利用已有标注数据; 该损失可根据细胞特征间的相似度大小自动调整损失权重, 使模型学习更有区分度的高维特征, 提高分类准确率. 实验采用 5 折交叉验证方法, 结果表明, 文中提出的尿脱落细胞分类方法在 Ideepwise 尿脱落细胞数据集上可达到 96.23%的敏感性和 97.2%特异性, 相比对比方法敏感性提升了 2.34 个百分点, 特异性提升了 4.6 个百分点.Abstract: In response to the challenges faced by current cell classification methods, including susceptibility to image quality variations and low classification accuracy due to limited labeled data, a cell classification method is proposed in this paper. First, the multi-level feature fusion module fuses semantic feature information and spatial feature information. Secondly, the adaptive comparison loss function can automatically adjust the loss weight according to the similarity between cell features, so that the model can learn more discriminative high-dimensional features and improve the classification accuracy. The experiments were conducted using 5-fold cross-validation. The results show that the urine exfoliated cell classification method proposed in this paper has 96.23% sensitivity and 97.2% specificity on theIdeepwise urine exfoliated cell dataset, which has a 2.34 percent improvement in sensitivity and 4.6 percent improvement in specificity compared with the existing methods.