Is there any detection model trained with negative samples? How to calculate the anchors shapes targeting nodules < 3 and >=3 mm ?

[abhishekmarkad]

I am using the script/config from https://monai.io/model-zoo.html to train retinanet detector,
While we put 10% negative samples in training data we had some problems. The Monai forum had a solution,followed it and made the changes,and were able to start training with custom data. Did experiments.Each experiment differs in terms of the nature of training data.
Changes done to include Negative samples in training

• Reference : https://github.com/Project-MONAI/tutorials/pull/1256/
• Added “StandardizedEmptyBoxd” transform in train.preprocessing_trasform in config/train.json obtained from https://monai.io/model-zoo.html ( Lung nodule CT detection )
  

Experiments

• Experiment 1 (positive + negative samples)

  • Nature of Training data : scans with nodules ranging from 2 mm onwards and 10% negative samples
  • Problem : The validation accuracy was zero till 115 epochs.
    

• Experiment 2 (positive + negative samples)

  • Nature of Training data : scans with nodules >=3 mm and negative samples
  • Problem : The validation accuracy was zero .
    

• I observe following output on console with Experiment 1 and 2.
  

• Experiment 3 (positive samples)

  • Nature of Training data : positive samples with nodules <3 mm & >=3mm
  • Problem : The validation accuracy increasing after 80+ epochs.
    

• Experiment 4 (positive samples)

  • Nature of Training data : Positive samples with nodules >=3 mm
  • Everything looks fine
    

Questions

1. Is it having anything with the size of nodules,I am wondering do i need to change the base anchors shape in config/train.json ?
2. Am I doing anything wrong while including negative samples ? Is there any model trained with negative samples ?
   Any help will be appriciated.

[abhishekmarkad]

Hello Monai Team, any thought ?