train slice - Train classification CNN using 2D slices

This option allows training a network on 2D slices. For more information on slices please refer to tensor extraction. There is no network type choice for slice as the only network type is the single-CNN.

One architecture is implemented in clinicadl for the slice mode: resnet18. If this architecture is chosen, the network is automatically initialized with the weights of a ResNet-18 trained on ImageNet.

Adding a custom architecture

It is possible to add a custom architecture and train it with clinicadl. Detailed instructions can be found here.

Introduction

The objective of this unique CNN is to learn to predict labels associated to images. The set of images used corresponds to all the possible slice locations in MR volumes. Slices at the beginning or at the end of the volume may be excluded using the discarded_slices argument.

The output of the CNN is a vector of size equal to the number of classes in this dataset. This vector can be preprocessed by the softmax function to produce a probability for each class. During training, the CNN is optimized according to the cross-entropy loss, which becomes null for a subset of images if the CNN outputs 100% probability for the true class of each image of the subset.

Running the task

Here is the command line to train a CNN on t1-linear outputs with the predefined architecture of ClinicaDL:

clinicadl train slice <caps_directory> t1-linear <tsv_path> <output_directory> resnet18

where mandatory arguments are:

• caps_directory (str) is the input folder containing the neuroimaging data in a CAPS hierarchy.
• tsv_path (str) is the input folder of a TSV file tree generated by clinicadl tsvtool {split|kfold}.
• output_directory (str) is the folder where the results are stored.

Common options

Options that are common to all train input and network types can be found in the introduction of clinicadl train.

The options specific to this task are the following:

• --slice_direction (int) axis along which the MR volume is sliced. Default: 0.
  • 0 corresponds to the sagittal plane,
  • 1 corresponds to the coronal plane,
  • 2 corresponds to the axial plane.
• --discarded_slices (list of int) number of slices discarded from respectively the beginning and the end of the MRI volume. If only one argument is given, it will be used for both sides. Default: 20.
• --use_extracted_slices (bool) if this flag is given, the outputs of clinicadl extract are used. Otherwise, the whole 3D MR volumes are loaded and slices are extracted on-the-fly.
• --selection_threshold (float) threshold on the balanced accuracies to compute the image-level performance. Slices are selected if their balanced accuracy is greater than the threshold. Default corresponds to no selection.

Outputs

The complete output file system is the following:

results
├── commandline.json
├── environment.txt
└── fold-0
    ├── cnn_classification
    │   ├── best_balanced_accuracy
    │   │   ├── train_image_level_metrics.tsv
    │   │   ├── train_image_level_prediction.tsv
    │   │   ├── train_slice_level_metrics.tsv
    │   │   ├── train_slice_level_prediction.tsv
    │   │   ├── validation_image_level_metrics.tsv
    │   │   ├── validation_image_level_prediction.tsv
    │   │   ├── validation_slice_level_metrics.tsv
    │   │   └── validation_slice_level_prediction.tsv
    │   └── best_loss
    │       └── ...
    ├── models
    │   ├── best_balanced_accuracy
    │   │   └── model_best.pth.tar
    │   └── best_loss
    │       └── model_best.pth.tar
    └── tensorboard_logs
         ├── train
         │    └── events.out.tfevents.XXXX
         └── validation
              └── events.out.tfevents.XXXX

Level of performance

The performance metrics are obtained at two different levels: slice-level and image-level. Slice-level performance corresponds to an evaluation in which all slices are considered to be independent. However it is not the case, and what is more interesting is the evaluation at the image-level, for which the predictions of slice-level were assembled.