SOCR ≫ DSPA ≫ DSPA2 Topics ≫

1 Deep learning Classification

  • Download the SOCR Alzheimer’s disease data
  • Preprocess the data and pool the MCI and AD cohorts (patients)
  • Build a multi-layer perceptron as a classifier (patients vs. controls) and select proper parameters
  • Classify AD and NC and report detailed evaluations, including cross table, accuracy, sensitivity, specificity, LOR, AUC
  • Provide some visualizations, e.g., histogram and model structure graph as we did in Chapter 14
  • Then, try to perform a multi-classes modeling (i.e., AD, NC and MCI) and report the classification results.

2 Deep learning Regression

  • Download the Allometric relationship data from the SOCR data archive
  • Preprocess the data and set density as outcome response feature
  • Create a MXNet feed-forward neural net model and properly specify the parameters
  • Train and predict the density using this model and report RMSE on the test data, evaluate the result and justify your evaluation
  • Output the model structure.

3 Image classification

Apply the deep learning neural network techniques to classify some images with pre-trained model as we did in Chapter 14:

4 (Extra Credit Problem) Deep Convolutional Networks for 3D Volume Segmentation

Use these 3D Brain Tumor Segmentation (BraTS) volumes for DCNN training and testing. Brain MR dataset contains \(257\) training images with corresponding labels and the dimensions of these MR images are 240*240 with 155 slices and 4 different imaging modalities including T1 (T1-weighted), T1C (contrast enhanced T1-weighted), T2 (T2-weighted), and FLAIR (Fluid Attenuation Inversion Recovery).

  • See this recent pub.
  • Design a clever 3D affine transformation mapping for volume augmentation (can use ITK affine transformation).
  • For pilot testing, subsample the data by \(5\) in each dimension, try to fit 3D volume in the existing TF/Keras/PyTorch tensor framework.
  • Configure and train the network, test with augmented volumes you generate.
SOCR Resource Visitor number Web Analytics SOCR Email