• 3D Segmentation & Classification with Keras
• Fine preprocessing with scikit-image
• Fine visualization for clarification
• Modified UNet for segmentation
• Modified VGG/Inception/ResNet/DenseNet for classification ensemble
• Fine hyperparameter tunning with both models and training process.

- config.py # good practice to centralize hyper parameters

- preprocess.py # Step 1, preprocess, store numpy/meta 'cache' at ./preprocess/

- train_segmentation.py # Step 2, segmentation with UNet Model
- model_UNet.py # UNet model definition

- train_classificaion.py # Step 3, classificaiton with VGG/Inception/ResNet/DenseNet
- model_VGG.py # VGG model definition
- model_Inception.py # Inception model definition
- model_ResNet.py # ResNet model definition
- model_DenseNet.py # DenseNet model definition

- generators.py # generator for segmentation & classificaiton models
- visual_utils.py # 3D visual tools

- dataset/ # dataset, changed in config.py
- preprocess/ # 'cache' preprocessed numpy/meta data, changed in config.py

- train_ipynbs # training process notebooks

• use SimpleITK to read CT files, process, and store into cache with numpy arrays
• process with scikit-image lib, try lots of parameters for best cutting
  • binarized
  • clear-board
  • label
  • regions
  • closing
  • dilation
• collect all meta information(seriesuid, shape, file_path, origin, spacing, coordinates, cover_ratio, etc.) and store in ONE cache file for fast training init.
• see preprocessing in /train_ipynbs/preprocess.ipynb file

Distribution of the lung part takes on a whole CT.

Tumor size distribution

• A simplified and full UNet both tested.
• dice_coef_loss as loss function.
• Periodically evaluate model with lots of metrics, which helps a lot to understand the model.
• 30% of negative sample, which has no tumor, for generalization.
• Due to memory limitation, 16 batch size used.

• A simplified and full VGG model both tested. Use simplified VGG as baseline.

Pictures tells that: hyperparameter tunning really matters.

• A simplified Inception-module based network, with each block has 4-5 different type of conv.
  • 1*1*1 depth-size seperable conv
  • 1*1*1 depth-size seperable conv, then 3*3*3 conv_bn_relu
  • 1*1*1 depth-size seperable conv, then 2 3*3*3 conv_bn_relu
  • AveragePooling3D, then 1*1*1 depth-size seperable conv
  • (optional in config) 1*1*1 depth-size seperable conv, and (5, 1, 1), (1, 5, 1), (1, 1, 5) spatial separable convolution
  • Concatenate above.

• use bottleneck block instead of basic_block for implementation.
• A bottleneck residual block consists of:
  • (1, 1, 1) conv_bn_relu
  • (3, 3, 3) conv_bn_relu
  • (1, 1, 1) conv_bn_relu
  • (optional in config) kernel_size=(3, 3, 3), strides=(2, 2, 2) conv_bn_relu for compression.
  • Add(not Concatenate) with input
• Leave RESNET_BLOCKS as config to tune

• DenseNet draws tons of experience from origin paper. https://arxiv.org/abs/1608.06993
  • 3 dense_block with 5 bn_relu_conv layers according to paper.
  • transition_block after every dense_block, expcet the last one.
  • Optional config for DenseNet-BC(paper called it): 1*1*1 depth-size seperable conv, and transition_block compression.

• Learning rate: 3e-5 works well for UNet, 1e-4 works well for classification models.
• Due to memory limitation, 16 batch size used.
• Data Augumentation: shift, rotate, etc.
• Visualization cannot be more important!!!
• coord(x, y, z) accord to (width, height, depth), naughty bugs.
• Put all config in one file save tons of time. Make everything clean and tidy
• Disk read is bottle neck. Read from SSD.
• Different runs has different running log dirs, for better TensorBoard visualization. Make it like /train_logs/<model-name>-run-<hour>-<minute>.
• Lots of debug options in config file.
• 4 times probability strengthened for tumors < 10mm, 3 for tumor > 10mm and < 30mm, keep for > 30mm. Give more focus on small tumors, like below.