Lung nodule detection multiclass training from scratch

[junxiant]

Hello, I've been trying to re-train the detection mode from scratch, but instead of doing detection for just "nodule" class, I have modified the dataset_fold0.json file, specifically the "label", by using the labelling from LIDC-IDRI dataset, at pylidc https://pylidc.github.io/annotation.html . Basically matching the json file "image" value and pylidc seriesinstanceuid.

So now there are a total of 5 classes, 0,1,2,3,4 .

I have modified the config file fg_labels to [0,1,2,3,4].

I have modified the luna16_training.py at line 245

before
    coco_metric = COCOMetric(classes=["nodule"], iou_list=[0.1], max_detection=[100])
after
    coco_metric = COCOMetric(classes=["Non_Solid_GGO","Non_Solid_Mixed","Part_Solid_Mixed","Solid_Mixed","Solid"], iou_list=[0.1], max_detection=[100])

But after the second validation after epoch 10 the val metrics are showing all 0.

Is there something i need to modify to make this work for multiclass? Or would i need to wait for a few more validation to confirm that it is not training properly

Here are some tf logs graphs

[KumoLiu]

Hi @Can-Zhao, can you help share some comments on this one? Thanks!

[Can-Zhao]

Thank you for reaching out.

• The coco metric may be very small in the first a few epochs. You may wait for 50 epochs to see.
• You maybe add --verbose flag. Details about matched anchors during training will be printed out. You can see if the hyper-parameters you set is correct.
• Have you looked at the visualized image and box in the tensorboard? Or maybe you could try a small lr and see how it goes?

[junxiant]

Thank you for reaching out.

• The coco metric may be very small in the first a few epochs. You may wait for 50 epochs to see.
• You maybe add --verbose flag. Details about matched anchors during training will be printed out. You can see if the hyper-parameters you set is correct.
• Have you looked at the visualized image and box in the tensorboard? Or maybe you could try a small lr and see how it goes?

Hello, u are right, it needed more rounds of validation. I was able to train it fully with the modified dataset labels, but it gave really poor results when i ran the evaluation script. i wonder if it makes sense to use the same evaluation script? since the original was for one "nodule" class but i have five classes now. mAR seems great, but mAP is very low

[Can-Zhao]

The evaluation script provided by LUNA challenge was not designed for multi-class detection. I have never tested it for multi-class detection. I would feel reporting Coco metric should be enough. Or you may run the evaluation script for each class separately.

[junxiant]

yes the coco metrics should be enough. the evaluation script does not work for multi class.

thanks for the help~

[junxiant]

oh just one more question

1. are there any plans to integrate more detection architectures into monai?