octavifs / lucanode Goto Github PK

Using the luna candidate annotations, implement and train a resnet and use that as a classifier, instead of the hand picked features. Compare the results with both approaches.

keras model.fit_generator returns a History object. This object stores the loss function and metrics for both the training and validation on each epoch. Save this output as a csv when storing the model (as .h5)

This is going to be useful to plot the evolution of the loss function over time and spot overfitting.

Average the weights of the cross validation folds trained by train_segmentation.py
The output of this average is then used in evaluate_scan_segmentation.py.

Currently scan evaluation relies on them being preprocessed into a 1x1x1mm spacing (this is the spacing used in the training of the network). Add an automatic resizing and cropping step so that evaluate_scan_segmentation.py can deal with any kind of CT as input

Currently the preprocessing for the nodule detection assumes that the lungs will have already been segmented. This is the case for LUNA (as the masks are available) but not so for other datasets.

If we want to extend the usefulness of lucanode to other datasets, it should be able to either to automatically segment lungs.

To this effect, I will employ the same UNET and lung masks already available to train another segmentation neural network. The output of this NN will then be used to perform the nodule segmentation.

Currently all the scans I'm using on evaluation are stored in .nii.gzformat. Make it so that any format supported by SimpleITK can be used. Also make it work with images in DICOM.

Hi, octavifs
How to cite your work? Which paper should be cite?
Thank you very much!
Best regards
Gu Yu

When training the UNET, the DICE score displays signs of overfitting after 5-7 epochs.

Ideas that might help overcome that:

• change training data on epoch end (randomize a percentage of it)
• add more and random transformations. Rotations in more angles, shearing, etc.
• purposely mislabel a low fraction (0.1) of the dataset on each epoch (by Gabriel)

Currently I'm manually extracting 6 features out of my labeled candidates. These are the base of my classifier for the FP reduction phase.

It would be good to test if using radiomics for feature extraction yields a better classifier than those hand picked features.