Text mining mri

• Description
• Project Structure
• Getting started
  • Container
    • Usage
  • Local
    • Usage
• References

Description

This is a machine learning project on report structuring [1] and grade classification. The goals are:

• to extract textual clinical entities from medical reports with the aim of structuring them with a template;
• to determine if a medical report contains evidences of osteoarthritis and to what grade.

For both tasks, the results were obtained by doing transfer learning of BERT [2] based models, fine-tuning those made available by Hugging Face.

The structuring of the reports was achieved by classifying the tokenized text entities, using BERT for the Named Entity Recognition. Instead, the classification of the degree of osteoarthritis was performed by classifying the tokenized sentences of the text of the reports using BERT for the Sequence Classification.

Project Structure

The src folder contains the core of the project, while the stack folder contains useful files such as the Dockerfile and the docker-compose.yml file. The Makefile allows to quickly run the project using the containerized and orchestrated software.

Below is the organization of the source project:

• The data folder has:
  • the ner folder with the manually annotated data for Named Entity Recognition;
  • the seq folder with the manually annotated data for Sequence Classification;
  • the tok folder with the generated reports used to build the tokenizers' vocabularies and to possibly train BERT from scratch for Masked Language Modeling;
• The data_utils folder has:
  • the data_processing.py script with some functions to process and prepare data (mostly for the tokenizer, BERT for MLM and annotation files);
• The model folder has:
  • a folder to store the experiments confusion matrices;
  • a folder to store the experiments train and test results;
  • a folder to store the experiments weights;
• The tasks folder has:
  • the masked_language_modeling folder with a script to process data and a script to do training and testing;
  • the named_entity_recognition folder with a script to process data and a script to do training, testing and inference;
  • the sequence_classification folder with a script to process data and a script to do training, testing and inference;
  • the tokenization folder with a script to build tokenizers;
  • the results_utils.py script with functions to plot and save experiments results;
• The main folder has:
  • the config.py script to set up useful information such as paths and training parameters;
  • the main.py script to run the training, validation, testing and inference;
  • the demo.py script to run a streamlit-based demo web application;

Getting started

Before starting, you can edit the config.py file and change the paths and folders names or training parameters. If set to something, the EXPERIMENT_ID parameter allows to create a folder inside the model one and keep the runs results separate from each other.

You can also edit the main.py script to select what to do between training, validation, testing and inference.

Container

To get started place in the base project folder and just type:

make build  # to build all docker containers

make shell  # to open a shell in the container

Usage

Container usage involves the same steps as local usage after entering the shell.

Local

Make sure to have python 3.10 installed.

Install the project requirements. Place in the src folder and type:

pip install -r requirements.txt

Usage

Place in the src folder.

If you want to try BERT-based models that involve building the tokenizer vocabulary (tok_train: True for bert_config in config.py) and training BERT for the MLM (mlm_train: True for bert_config in config.py), the first time it is launched, it is also necessary to generate the texts of the reports to be used as a dataset, therefore it is necessary to run:

python data_utils/data_processing.py

Otherwise, and in any case even after following the previous step, to run the main.py script and start the training/validation/testing/inference, type:

python main.py

To run the demo web application, type:

streamlit run demo.py

Before launching the demo make sure to have generated the neural network weights and have the config.py parameters set accordingly.

The demo web application will be available on localhost:8501.

References

[1] Kento Sugimoto et al. (2021). Extracting clinical terms from radiology reports with deep learning https://www.sciencedirect.com/science/article/pii/S1532046421000587

[2] Jacob Devlin et al. (2019) BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding https://arxiv.org/pdf/1810.04805.pdf