A Fusion NLP Model for the Inference of Standardized Thyroid Nodule Malignancy Scores from Radiology Report Text

This repository contains the workflow of our proposed algorithm to infer TI-RADS scoring from clinical notes. Based on a given csv file with TI-RADS reports, our solution does:

1) Segment all nodules in each report
2) Classify each nodule
3) Save the results (Accurary, F1, and Confusion Matrix) to a file

Publication

Accepted to AMIA(American Medical Informatics Association) 2021 ANNUAL SYMPOSIUM

Pipeline

Our pipilne representation is illustraded bellow:

Fusion-BERT Classification System

Our proposed model arquitecture is illustraded bellow:

The output probabilities of the three independent models and concate-nated and fed into a logistic regression meta-learner.

Interface Design

A friendly user interface was developed, using python and PyQt. The interface can handle a single nodule classification as well as prediction over all reports on a file. For a single nodule prediction, the user can input a nodule description and the application will predict the output with the highest probability. In this module, the application also gives the probability of all categories, giving a better indication and explanation to the user. The user can also load a file with multiple reports and perform a classification on all nodules at once. Once the file is loaded, our application automatic segment every report to extract all nodules.

A runing example of the GUI application from a report and also from a single nodule:

Application Usage:

cd GUI/
python3 application.py 

Terminal Usage:

python3 main.py -inputfile myfile.csv -column_name report -output_file results.txt

Usage: main.py 	[-inputfile INPUTFILE] - Choose the input CSV file; Default: indiana.csv
				[-column_name COLUMN_NAME] - Provide the column name for the report text; Default: report
           		[-output_file OUTPUT_FILE] - Choose an output file name; Default: results.txt
           		[-remove_structured True/False] - Choose to remove or not structured reports - If false, only nodules with free-text form of description will be used; Default: True

Install Dependencies:

apt-get update && apt-get install -y python3 \ python3-pip

pip3 install --upgrade pip

pip3 install -r requirements.txt

Install TIRADS-Inference

cd

git clone https://github.com/thiagosantos1/TIRADS-Inference.git

Download Our Pre-trained model

Open the following link and download the model

https://drive.google.com/file/d/1qhkilrbjO_heIUMvzCdRdEUcdXjRJVpQ/view?usp=sharing

Unzip and move all files to TIRADS-Inference/BERT/

cd Downloads(Or folder where file were downloaded)

unzip BERT_model.zip -d BERT_model

mv BERT_model/* TIRADS-Inference/BERT/

Use model to predict TI-RADS

cd

cd TIRADS-Shared

python3 main.py [usage]

Contributors

Phd. Thiago Santos

Dr. Imon Banerjee

Dr. Judy Wawira

Dr. Omar Kalls