solution.txt

Won't go in too much detail here, but here's the general pipeline of my current LB submission (~0.42):

Pre-process all of the stage 1 and LUNA16 images into a common format, segmenting the lungs and rescaling to 1x1x1mm voxels;
Extract 64x64x64 chunks from all entries on the LUNA16 annotations_excluded.csv file, and a random sample of the non-nodule entries from candidates.csv as a negative reference;
Train a 3D VGG derivative on augmented entries from the features above with the labels being classes based on the size of the nodules;
Chop up each scan from the stage 1 dataset into overlapping chunks to fit the above network, capturing the output label and the final convolution step;
Build a feature set from the output above, aggregating over all chunks;
Train a XGB classifier with the features and the stage 1 labels;
Evaluate the classifier on the test set;
The devil is in the details, and all of this takes considerable time to fit all together, but I believe the other top scores right now are using similar techniques, basically the idea of training on LUNA16 data and then reusing that to extract features from the competition dataset.

I think working with 3D ConvNets here is the way to go, and I don't think reusing trained VGG16 or ResNet weights will go much better than the ~0.5 of most approaches using it right now. I think the main problem is having at most 3 layers at once, and the original training having no spatial correlation over the (originally color) layers.

A design based on VGG or ResNet with a few more layers may work decently enough, but you still need to train it up with something else than the competition data if you want to get anywhere.

Good luck and have fun :-)