ec601-project's Introduction

ec601-project RSNA Pneumonia Detection

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

Author: Min Zhou ([email protected]), Andrew Stoycos ([email protected])

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:

Medical image analysis and processing.
A machine learning model that can predict pneumonia and return the result.
- Pneumonia Positive
- Pneumonia Negative
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.

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.

ec601-project's People

Contributors

Stargazers

ec601-project's Issues

For your README file, It is good to provide a configuration guidance for a beginner user just like me. But when following the instructions, I faced some problems, you may want to modify corresponding parts at your README file:

Some incorrect file names:

I think it is ec601-project.

For me, the folder name I downloaded is called "all" instead of input. I need to do the modification to avoid the run time error.

You may also consider to add the instruction of installing Jupyter Notebook before step3, I face the "Invalid requirement: 'jupyter notebook'" error when doing step 3, so I can only add those packages manually.
My operating system is Windows, and I also need a configuration of the Jupyter‘s ’work path before run it, Not sure if it just happens to me.

Thanks for the guidance anyway.

Package Import Missing?

There is a run time error:

I guess it is because the lacking of sklearn package?

Model selection

Hi,

I notice that code about the mask-rcnn model has been uploaded. Actually, yolo and mask-rcnn are both excellent models for object detection. As for me, I don't think there will be much difference for your task. I believe the training process is more crucial here. As you can see in your current result, the key point is that you should use more data for training. Then, you may still need much time for adjusting parameters to make it best.
Are you still trying to work on these two models separately? Considering the time and effort, I suggest it is better to focus on one model and keep training. Thanks!

minzhou1003 / ec601-project Goto Github PK