Optimize 2D-CNN

[ValleSoYeah]

• [x] implement tracking of model performances

[deneick]

changes in preprocessing

• [x] use images without skull
• [ ] try using few multiple slices (treat same as channels)
• [x] train same model on different sets of slices to identify most useful slices ( => more or less the slices that were provided as gifs, bad performing model did not allow clear judging of the performance of different layers
• [x] use multiple adjacent slices for training
• [x] try with rescaling to max intensity

optimize to the model

• [ ] try using cropped regions (cropping layer?)
• [x] use a filter to identify important regions (attention layer? filter of fulll-image-size? Locally connectedConv2D layer!)
• [x] try adding augmentation layers (very slow since GPU can't be used, conflicts with LocallyConnectedConv2D => only use flip, if any)
• [x] Majority voting for models trained from different viewing angles
• [ ] Cyclic learning rate
• [x] Batch normalization (adds some Regularization)
• [x] Using pca to get eigenbrains
• [x] Resnet50 or vgg16 with random weights (didn't work, predicts only one class)
• [x] Transfer learning (works but poor performance)

changes to the training-set

• [x] remove young patients to deal with large young-vs-old differences
• [ ] undersampling to deal with data imbalance (remove only young to fight both issues)

[ValleSoYeah]

Recent sucesses:

• Locally connected conv2d improves validation accuracy and reduces overfitting even in very simple model
• 98% validation accuracy with pretrained VGG16 + LocConnectedConv2D layer (without dropping data and loading multiple slices for training)
• still only 73% test accuracy in best model => only single slices used for X_test (mimics application of the model to predict from 2D images alone)

[ValleSoYeah]

Try next:

• [x] Shift each ADNI brain so its mean aligns with the mean of all OASIS brains
• [x] use VGG16_LocConn model to identify best slices
• [ ] modify get_slice functions to allow loading of thin 3D images (3 layers, but connected like in an RGB image)
• [ ] can eigenbrains increase training time? do CNNs work with eigenbrains?

[deneick]

• I tried using pca with svm with 2d and 3d. Problem: we can only have as many components as data samples.

• I changed the loading of the data, so it loads the data as uint8 and int16 instead of converting to float64 Makes the loading of the data faster and needs much less memory, but when applying pca or cnn, the data needs to be converted to float32 or float64 Data also gets converted when normalizing in current code

• data can also get sliced before even loading the array. Could make loading slices faster nibabel doc

• cnns can be built with multiple inputs. Can be used for different views code