Identification of "AIA terminal nodule" and "AIA central nodule" using DenseNetFCN-3D.

Based on two different 3D movies (20190805_SJR3.2.2_w1_s2.nd2 and 20190805_322_w1_s2.nd2), provided by the Laboratory of the Physics of Biological Systems, our task is to identify and distinguish two parts (the terminal nodule and the central nodule) of a neuron type called AIA.
For our work we considered all frames as different volumes, not taking into account time continuity.
We used a DenseNetFCN-3D with 4 dense blocks of 3 layers per block (mostly because of computation constraints).

Deliverables:

• DataGenerator.py python file: Create generator for training and validation, user may apply image augmentation on the fly.
• DenseNet3D.py python file: Implementation of the DenseNetFCN-3D from GalDude33.
• generateFramesMasksFromVideo.ipynb notebook: Generate frames/masks with desired shape from the video/ground_truth files.
• Training.ipynb notebook: Feed the DenseNetFCN-3D with frames/masks pairs.
• VisualizationPrediction.ipynb notebook: Predict an example mask with the DenseNetFCN-3D model.
• ConfusionMatrix.ipynb notebook: Compute the accuracy and the F1 score for the 'AIA terminal nodule' (class 1) , the 'AIA central nodule'(class 2) and the other cells (class 0).

(keep all the files in the same directory)

How to train the DenseNetFCN-3D, predict masks and note the model:

We executed the code on Google Colab by following these steps:

1. The training samples and their corresponding masks need to be stored as .npy files. Their names should be frame_i.npyand mask_i.npy respectively (for i going from 0 → maxTime) and frames/masks need to be stored in two different folders. If you gives the good PATHs in the second cell of generateFramesMasksFromVideo.ipynb, this notebook does this steps for you.
   P.S: To help our model, we labeled ALL cell parts that were not AIA central/terminal noduls with the same value. Thus, the model only has to distinguish 3 types of cells.
2. Give the wanted paths and variables values to all the second cell's variables of Training.ipynb notebook. Then execute all cells. It will train a DenseNetFCN-3D and save the model.
3. Once the model has been trained, it is possible to have a visual idea of the prediction with the VisualizationPrediction.ipynb notebook. First, change the variables/paths of the second cell and then run all the cells. The last cell shows the image, the mask and finally the predicted mask.
4. Finally, the ConfusionMatrix.ipynb notebook can be used (once the PATHs from cell 2 have been changed) to compute the accuracy and the F1 score for the 'AIA terminal nodule' (class 1) , the 'AIA central nodule'(class 2) and the other cells (class 0).

REFERENCES: DenseNet3D from GalDude33

Deadline: December 19th 2019, 23:59