PNEUMONIA DETECTION USING CNN IN FEDERATED LEARNING ENVIRONMENT

Abstract

As the symptoms of certain fatal diseases are not readily visible on CT as well as X-ray images, computerized identification of serious respiratory conditions like pneumonia, caused by inflammation of the air sacs of the lung alveoli, poses a substantial problem in medical imaging. Moreover, it is a crucial job because pneumonia affects millions of people each year. The main purpose of this work is to develop an innovative machine learning technique in federated learning environment using a deep learning neural network to solve the aforementioned problem. A convolutional neural network is used for detection of the disease. The proposed technique was trained and tested using a collection of 5856 images with tags from one of Kaggle's medical imaging challenges. Anterior-posterior chest X-ray pictures were obtained. Federated Learning is a collaborative learning unlike the traditional machine learning where the entire learning happens in a centralized server .Many organizations like hospitals are reluctant to share their patient’s data due to privacy concerns. Federated Learning comes as a solution to this problem. In FL method, models are trained locally on the client side. Data is not transferred outside the organization. Thus FL helps the model to attain finer accuracy by training model on vast amount of data without sharing. Thus this method of training ensures data privacy and security. In this work, pneumonia detection using CNN in a federated learning environment has been developed. This model uses Federated averaging as the aggregation mechanism. Flower framework is used for the implementation of federated learning environment. An accuracy of 97 percent is obtained for 10 communication rounds which is at par with the accuracy obtained with centralized learning implementation.

Specification

Description:• Federated Averaging (FedAvg) , Federated Stochastic Gradient Descent (FedSGD) are some of the most common algorithms used in Federated Learning Techniques .
• Other algorithms include scaffold, FedXpro, FedBn etc . Federated learning thus enable us to overcome various issues associated with typical machine and deep learning models where data privacy is a major concern. The main idea behind federated learning is decentralized data.
• The interesting fact is that the data that is used to train the model in one client, never leaves that client or organization. The weights, biases and other parameters learned by that individual model leaves that client.
• To implement Pneumonia detection model in Federated Learning Environment, first the medical imaging data (such as X-rays or CT scans) from different sources is to be gathered and pre-processed. Then, use a convolutional neural network (CNN) architecture for image classification. The model parameters would be trained locally on each participant organization or hospital's data, and the updates would be aggregated and leveraged to update the global model.
• TOOLS USED
Language: Python
Documentation: Microsoft Word, Google docs, LaTex
Software: Google Colab,Tensorflow 2.x, Pytorch, Flower framework
Hardware: Intel® Core™ Processor i5-7660 U CPU @ 2.50GHz;
Installed 4.0 GB RAM; Google Colab RAM 80GB
64-bit Operating System, x64- based processor.
, Claims:1. We claim that a CNN model is suggested as a practical and accurate approach to the X-ray based pneumonia detection problem. The insertion of a dropout layer among the network's convolutional layers was the primary innovation.
2. We claim that the built a CNN model from scratch as opposed to using transfer learning to deploy pretrained networks. The suggested model outperforms its counter alternatives in experiments in terms of efficiency and accuracy, reaching recall and precision well above 97 percent with predictions made in just milliseconds.
3. We Claim that the model has good durability along with accuracy. If the model used with proper system the mortality rate due to Pneumonia can be reduced.
4. A deep learning model is developed to detect pneumonia using dataset containing chest X rays(CXR images) using Convolutional Neural Networks. Thereby reduce mortality rate and intensive health care requirement.
5. We claim that the A Federated Learning Pneumonia Detection model using deep learning simulation is developed using flower framework.
6. We claim that the Federated learning concept is introduced to this model to address data privacy issues arising due to sharing of information of patients.
7. We claim that through collaborated learning approach, that is using Federated Learning environment a more efficient model is developed.
8. We claim that other datasets can also be tested and trained for the generalization of the model. Methods are to be devised to address privacy issues.
9. Methods are to be researched to reduce the communication time required for training. Data heterogeneity and imbalance limit constraints are to be efficiently addressed and how well models can be aggregated in another consideration for further study. Future research will focus on optimizing aggregation process while taking privacy costs into account.

Documents