Interpreting Raman spectroscopy towards diagnostic purposes : an explainable deep-learning based approach

Based on Raman spectra data, this project is part of a larger project to investigate new, less intrusive and faster diagnostic methods. The purpose of this part is to suggest an interpretation of the CNN learning mechanism proposed in this paper. This research hopes to implement a method in order to allow an intuitive understanding and further application.

In order to achieve this objective, this project investigates two different approaches, the second being a variant of the first:

1. Class Activation Mapping - CAM (the original method was defined in this paper and later applied to spectral data in this work)
2. Gradient Class Activation Mapping - Grad-CAM (this method is originally proposed here)

Technical requirements

This code was originally executed with 4 GPUs, on the Future SOC lab server. But, theorically, it can be run on different configuration using the librairies and the environment specified in the folder server-configs.

Organisation of the repository

This repository can be divided into 7 parts:

1. Scientific reports
2. Technical set-up
3. Transformation of the original model using the Pytorch library
4. Bacterial study case
5. Implementation of CAM method
6. Implementation of Grad-CAM method
7. Visualisation of the different results obtained

Scientific reports :

All files written during this project are grouped in the report folder. There are :

• Presentation_PDF_version : it is the presentation of this work in PDF.
• Presentation_PowerPoint_version : it is the original presentation of this work.
• Project_synthesis : it the synthesis of this project (5 pages).
• Report_complete : it is the complete report of this work.
• Report_short_version : the report of this work in shorter.

Technical set-up :

All the files needed to set up the environment required to run this work are contained in the server-configs folder. This folder contains several shell scripts :

• create_docker_container : script to automate the creation of the specified docker container.
• install_ssh : script created to able ssh into the docker container.
• update_environment : script to automate the update of the conda environment in the docker container.

In addition, it contains the jupy_config file which specifies the configuration of jupyter notebooks, and the subfolder pytorch_conda_config which contains :

• Dockerfile : the Dockerfile used to create the container.
• environment : contains the definition of the conda environment, with all the librairies needed to compute this project.
• docker_clean : script for clean the docker environment in order to re-build a container.
• condasetup : script to activate the conda environment needed.

Moreover, to run this code an implementation of the cubic spline is used. It can be found in the folder torchcubicspline and its taken from this GitHub repository.

Pytorch transformation :

The main code of this transformation is contained in scripts folder. It is divided in two main parts :

1. data_manager : it contains all the file needed to manage the COVID data.
   • Datasets : class to create a dataset composed of multiple Raman Dataset in order to compute the Leave-One-Patient-Out Cross-Validation.
   • RamanDataset : class to define a single Raman Dataset.
   • data_script : contains all the functions needed to read, split and manage the COVID data.
2. model_manager : it contains all the file needed to create and manage the CNN used in this project.
   • ConvNet : class to define the CNN in Pytorch.
   • model_script : contains all the functions needed to create, parallelize, train and test the model.

This transformation is prepared and tested in several Jupyter notebooks available in the Model_Pytorch folder :

1. Raman_model_Pytorch : This is the original notebook in which the code condensed in the scripts folder is developed.
2. Model_with_PD_Dataset : This notebook aims to test the developed model on Parkinson's data to confirm the results.
3. New_test_set : In this notebook, the idea is to separate a patient before performing the Leavo-One-Patient-Out Cross-Validation method in order to simulate a completely new patient for all models.

All models created during this phase are not available pre-trained in this repository for size limitationr reason, but can be obtained on reasonalble request at singlan.nina@gmail.com.

Bacterial study case :

The data and model used for the bacteria study case are taken from this work, and is available into the bacteria folder.

CAM :

The Class Activation Mapping is implemented and tested into the cam folder. The files concerning the implementation are :

• importance : which contains the code in order to obtain the importance for an input spectra.
• CAM : which contains the code needed to compute the Class Activation Mapping method on an entire dataset.
• CAM_on_bacteria : contains the computation of the method on the bacterial case study.
• CAM_on_covid : contains the computation of the method on the COVID case study.

As explained in the scientific reports, the Class Activation Mapping method need a specific architecture which is created and trained for both case study, thus, the CAM model for the bacterial dataset is in the bacteria folder. Then, it is defined in resnetCAM, trained and fine-tuned in CAM_Model and finally saved in CAMModel and CAMModelFine-Tuned. And, for the COVID case, the model is defined in the file ConvNetCAM in the scripts/model_manager folder. Then, is trained in the notebook CAM_Model located into the Model_Pytorch folder. Finally, each model generated by the Leave-One-Patient-Out Cross-Validation is saved into the saved_models/CAM_Covid folder.

Grad-CAM :

The Gradient Class Activation Mapping is implemented and tested into the cam folder. The files concerning the implementation are :

• Grad-CAM : which contains the code in order to get the Gradient and the Feature Maps for a model.
• compute_Grad_CAM : which contains the code needed to compute the Grad-CAM over a entire dataset.
• Grad-CAM_on_bacteria : contains the computation of the method on the bacterial case study.
• Grad-CAM_on_Covid : contains the computation of the method over the COVID dataset.

Visualisation :

Visualisation results are not available in this repository for size limitationr reason, but can be obtained on reasonalble request at singlan.nina@gmail.com.

It contains differents folders :

1. CAM_Results : in which all the results produced by the Class Activation Mapping can be found.
   • Bacteria : contains all the results produced on the bacterial case.
     • Each folder named by an integer contains the results obtained on the spectra belonging to the class identified by this integer.
       • Graphs : contains the visualisation of the results.
       • Values : contains text files in which the values visualised in the Graphs folder are listed.
     • Statistics : contains the statistics obtained.
       • BarChart : contains the visualisation of the statistics.
       • Values : contains text files in which the values visualised in the BarChart folder are listed.
   • COVID : contains all the results produced on the covid case.
     • Each folder named by "model i" where i is an integer contains the results obtained with the model trained on the i-th fold of the Leave-One-Patient-Out Cross-Validation.
       • Folder named by an integer contains the results obtained on the spectra belonging to the class identified by this integer.
         • Graphs : contains the visualisation of the results.
         • Values : contains text files in which the values visualised in the Graphs folder are listed.
       • Statistics : contains the statistics obtained.
         • BarChart : contains the visualisation of the statistics.
         • Values : contains text files in which the values visualised in the BarChart folder are listed.
2. Grad-CAM_Results :
   • Bacteria : contains all the results produced on the bacterial case.
     • Each folder named by an integer contains the results obtained on the spectra belonging to the class identified by this integer.
       • Graphs : contains the visualisation of the results.
       • Values : contains text files in which the values visualised in the Graphs folder are listed.
     • Statistics : contains the statistics obtained.
       • BarChart : contains the visualisation of the statistics.
       • Values : contains text files in which the values visualised in the BarChart folder are listed.
   • COVID : contains all the results produced on the covid case.
     • Each folder named by "model i" where i is an integer contains the results obtained with the model trained on the i-th fold of the Leave-One-Patient-Out Cross-Validation.
       • Folder named by an integer contains the results obtained on the spectra belonging to the class identified by this integer.
         • Graphs : contains the visualisation of the results.
         • Values : contains text files in which the values visualised in the Graphs folder are listed.
       • Statistics : contains the statistics obtained.
         • BarChart : contains the visualisation of the statistics.
         • Values : contains text files in which the values visualised in the BarChart folder are listed.
3. train_visualization : contains the visualisation of the training phase for each model created by the Leave-One-Patient-Out Cross-Validation.

In addition, in the folder cam, the two files named visualize_results and run_visualization are dedicated to the calculation of statistical results.