[MICCAI2019] Overfitting of neural nets under class imbalance: Analysis and improvements for segment

作者单位:
Zeju Li,Biomedical Image Analysis Group, Imperial College London


分析

在前景比较少(类别不平衡)的分割场景中,逐步减少训练数据量,作者发现precision稳定,sensitivity下降明显。
[MICCAI2019] Overfitting of neural nets under class imbalance: Analysis and improvements for segment
作者进一步可视化分类层的输出,发现1、前景离决策面太近,背景离决策面较远;2、随训练数据量减少,前景train和test的均值点gap变大,且test的均值点逐渐进入背景区域,说明随数据量减少,过拟合现象越严重
[MICCAI2019] Overfitting of neural nets under class imbalance: Analysis and improvements for segment
根据上述分析,作者希望通过显式地将前景的logits拉离决策面,来提升最终的分割精度。基于此,作者对传统loss引入了基于前景的非对称性。

改进

1、margin loss

[MICCAI2019] Overfitting of neural nets under class imbalance: Analysis and improvements for segment
改进为:
[MICCAI2019] Overfitting of neural nets under class imbalance: Analysis and improvements for segment

2、focal loss

[MICCAI2019] Overfitting of neural nets under class imbalance: Analysis and improvements for segment
改进为:
[MICCAI2019] Overfitting of neural nets under class imbalance: Analysis and improvements for segment

3、adversarial training

4、mixup

改进后结果

根据可视化结果,改进后虽然train和test的gap仍存在,但逐步将test的均值点移回到了前景区域。从表1来看,分割精度也确实有了较大提升。
[MICCAI2019] Overfitting of neural nets under class imbalance: Analysis and improvements for segment