Noisy Label Learning

Note

Some NLL methods only use noise-robust loss functions without complex training process, and just combining the standard SegmentationAgent with such loss function works for training.

NLL Configurations

In the configuration file for noisy label learning, in addition to those used in standard fully supervised learning, there is a noisy_label_learning section that is specifically designed for NLL methods. In that section, users need to specify the method_name and configurations related to the NLL method. supervise_type should be set as “noisy_label” in the dataset section.

For example, the correspoinding configuration for CoTeaching is:

[dataset]
...
supervise_type = noisy_label
...

[network]
...

[training]
...

[noisy_label_learning]
method_name  = CoTeaching
co_teaching_select_ratio  = 0.8
rampup_start = 1000
rampup_end   = 8000

[testing]
...

Note

The configuration items vary with different NLL methods. Please refer to the API of each built-in NLL method for details of the correspoinding configuration. See examples in PyMIC_examples/seg_nll/.

Built-in NLL Methods

Some NLL methods only use noise-robust loss functions. They are used with a standard fully supervised training paradigm. Just set supervise_type = fully_sup, and use loss_type to one of them in the configuration file:

• GCELoss: (NeurIPS 2018) Generalized cross entropy loss.

• MAELoss: (AAAI 2017) Mean Absolute Error loss.

• NRDiceLoss: (TMI 2020) Noise-robust Dice loss.

The other NLL methods are implemented in child classes of pymic.net_run.agent_seg.SegmentationAgent, and they are:

• CLSLSR: (MICCAI 2020) Confident learning with spatial label smoothing regularization.

• CoTeaching: (NeurIPS 2018) Co-teaching between two networks for learning from noisy labels.

• TriNet: (MICCAI 2020) Tri-network combined with sample selection.

• DAST: (JBHI 2022) Divergence-aware selective training.

Customized NLL Methods

PyMIC alo supports customized NLL methods by inheriting the SegmentationAgent class. You may only need to rewrite the training() method and reuse most part of the existing pipeline, such as data loading, validation and inference methods. For example:

from pymic.net_run.agent_seg import SegmentationAgent

class MyNLLMethod(SegmentationAgent):
  def __init__(self, config, stage = 'train'):
      super(MyNLLMethod, self).__init__(config, stage)
      ...

  def training(self):
      ...

agent = MyNLLMethod(config, stage)
agent.run()

You may need to check the source code of built-in NLL methods to be more familar with how to implement your own NLL method.

In addition, if you want to design a new noise-robust loss fucntion, just follow Fully Supervised Learning to impelement and use the customized loss.