ec601-project RSNA Pneumonia Detection

Kaggle link: https://www.kaggle.com/c/rsna-pneumonia-detection-challenge

Author: Min Zhou (minzhou@bu.edu), Andrew Stoycos (astoycos@bu.edu)

Web App

Poster

Project progress: Trello

File Instruction:

• working is the folder including a Jupyter Notebook for analyzing and processing our dataset.
• app folder is the web application of our product.
• yolo_model folder contains some model config files and a Jupyter Notebook used to train the model.
• requirements.txt contains some necessary python libraries to run our code.
• MASKrcnn_model contains the python code for Mask-RCNN model.
• Lung_Segmentation contains python code for segmenting lungs before training the Mask-RCNN model.

Product statement:

• Doctors, patients and medical professionals, need a product to help them to improve the efficiency and reach of diagnostic services.

• This is an AI (machine learning/deep learning) model which can automatically detect a signal of pneumonia in medical images.

• Current pneumonia diagnosis must be completed by a medical professional following a chest X-Ray and physical exam

• Our product will automate initial detection (imaging screening) of potential pneumonia cases and create bounding boxes around the areas of interest in order to prioritize and expedite their review.

MVP:

• [x] Medical image analysis and processing.
• [x] A machine learning model that can predict pneumonia and return the result.
  • Pneumonia Positive
  • Pneumonia Negative
  Normal Negative
  Non-Normal Negative(other disease)
  • [x] A website application to make the software accessible.

System Diagram

Data Source

• The RSNA training dataset is made up of 26684 chest radiographs in DICOM format can be download from kaggle.
• This is a binary class data: 68% pneumonia negative and 32% pneumonia negative with bounding box labels.
• The resolution of raw image is 1024x1024.

Machine Learning Model (Deep Learning)

YOLO v3 Model

YOLO is an open source real-time object detection model. It has 106 layers and it's using localization, classificaiton, regression and Focal loss. The benifits of using YOLO v3 are listing below:

• Due to computational limitation, YOLO is fast and accurate for our problem.
• Easy to change the achitecture and parameters.
• There are a lot of resources to train and optimize the model.

Mask-RCNN model

To test the data anlysis part

Installation:

• python 3
• Jupyter Notebook
• virtualenv
• Note: All the code is tested on Linux 4.9.0-5-amd64 and NVIDIA Tesla K80 GPU. If you are trying to run it on Windows, you may need to use virtual machine.

1. Download this repository:

git clone https://github.com/minzhou1003/ec601-project.git

2. Set up and activate virtualenv inside that folder.

cd ec601-project
virtualenv --python python3 env
source env/bin/activate

3. Install python libraries:

pip install -r requirements.txt

4. Download the dataset

Download the dataset in the same directory of this project. You should get a folder called input.

5. Run the Jupyter Notebook

Go to the working directory and open your jupyter notebook:

cd working
jupyter notebook

To run and check demo of our application:

See our app instruction.