A Distributed Privacy Preserved Federated Learning Approach for Revolutionizing Pneumonia Detection in Isolated Heterogenous Data Silos

International Journal of Mathematical, Engineering and Management Sciences Vol. 10, No. 5, 1324-1350, 2025 https://doi.org/10.33889/IJMEMS.2025.10.5.063 A Distributed Privacy Preserved Federated Learning Approach for Revolutionizing Pneumonia Detection in Isolated Heterogenous Data Silos Shagun Sharma Chitkara University Institute of Engineering and Technology, Chitkara University, Punjab, India. & School of Computing Science & Engineering, VIT Bhopal University, Sehore, Bhopal, Madhya Pradesh, India. E-mail: , Kalpna Guleria Chitkara University Institute of Engineering and Technology, Chitkara University, Punjab, India. Corresponding author: , (Received on February 4, 2025; Revised on April 15, 2025; Accepted on April 29, 2025) Abstract Pneumonia is a respiratory lung contamination that ranges in severity from mild to lethal outcomes. The analysis of tomographic images is the most significant method of pneumonia detection. The image analysis requires expertise and proficiency to diagnose the disease correctly. The medical reports with multiple diseases have overlapping symptoms, which may lead to misdiagnosis and deferred identification. The misdiagnosis results in increased healthcare costs, worsened medical conditions, and legal implications. Centralized deep learning enhances the feature extraction process and optimally improves the prediction outcomes; however, these models have data privacy concerns due to centralized storage systems. The healthcare departments follow the Health Insurance Portability and Accountability Act. (HIPAA) to maintain the retaining of patient data and improve the portability and continuity of health insurance coverage. In the proposed work, federated learning has been utilized to enhance data privacy and deal with imbalanced and diverse data silos. This distributed privacy-preserved model has been employed with a pooled dataset curated from multiple sources in a 5-client architecture. The model was implemented with the FedAVG aggregation technique in independent and identically distributed (IID) and non-IID data distributions. The outcomes of the model exhibit 87.62% accuracy with IID and 86.15% accuracy with non-IID distributions. The comparison of these outcomes with the existing studies shows that the proposed model outperforms by exhibiting better performance and resulting in the minimum loss of 0.4041 and 0.4139 with IID and nonIID distributions, respectively. Keywords- Deep learning, InceptionV3, Distributed architecture, Federated learning, Pneumonia detection, Centralized learning. 1. Introduction In 2019, the number of Covid-19 cases started to increase, resulting in a global health crisis and widespread transmission, hence WHO declaring it as a pandemic in March 2020 (Kafadar et al., 2022). It caused a rapid increase in mortality rates due to the unavailability of vaccines and medications. This pandemic directly impacted the patient care system and economic growth that limited medical resources and healthcare equipment’s worldwide (Holt et al., 2020). The sudden surge of Covid-19 cases and immediate shutdowns led to a decrease in the available healthcare assets and medicine supplies. Covid-19 affects the respiratory system and results in reducing surfactant secretion (Alipoor et al., 2020). The decrease in secretion causes the alveoli to collapse, resulting in pneumonia. In the past 20 years, the mortality cases caused by pediatric pneumonia have significantly decreased (Kanwal et al., 2024). However, these cases started to increase during the pandemic, when the available healthcare resources were insufficient for the treatment of Covid19. Pneumonia is respiratory lung contamination caused by bacteria and viruses and ranges from mild to 1324 | https://www.ijmems.in Sharma & Guleria: A Distributed Privacy Preserved Federated Learning Approach for … lethal outcomes (Adjei-Mensah et al., 2024). The advancement in pneumonia stages may affect other organs, such as the nervous system, heart, and lungs, along with blood vessels (Hatmi, 2021). The identification of pneumonia is challenging to the lower income countries due to the complex structure of the disease (Kundu et al., 2021). The other problem associated with this disease is its misdiagnosis due to the similar kinds of symptoms in tuberculosis, Covid-19, and lung cancer. As per the WHO report, pneumonia affects people over the age of 65 and patients with a weakened immune system the most (Pneumonia, n.d.). There are various other kinds of respiratory diseases, such as silicosis, chronic bronchitis, cystic fibrosis, asthma, lung cancer, and tuberculosis. The statistics presented in Goyal & Singh (2023), show that lung cancer has affected 8 million people to date; however, this patient count is less if compared to the 15-month cases of Covid-19 and pneumonia. Respiratory infections and diseases are detected using the tomography imaging process, specifically with X-ray scans; however, it requires expertise and proficiency to correctly diagnose the disease. Patients with multiple diseases may have overlapping symptoms in their reports, which creates complexity in accurately diagnosing the associated disease, and the obliviousness of the medical imaging process leads to misdiagnosis of the disease. The misdiagnosis of the disease may result in worse medical conditions, increased healthcare costs, psychological impact, and legal implications, along with lethal results. There are various existing studies that have used different kinds of artificial intelligence (AI) tools on medical images to accurately identify diseases. The deep learning (DL) models are also used for Alzheimer's, Parkinson’s, pneumonia, cancer, and Covid-19 disease detection (Hariri & Avşar, 2023; Kafadar et al., 2022; Kundu et al., 2021; Prakash et al., 2023b; Prakash et al., 2023c; Rahimzadeh & Attar, 2020). These models are highly efficient and optimal for providing improved performance of disease detection by performing proficient image processing and feature extraction. The Xception, InceptionV3, VGG16, VGG19, ResNet50, CNN, and ensemble models have been used by various researchers to predict pneumonia from lung CXRs (Alyasseri et al., 2022; Gulati et al., 2022; Hasan et al., 2021; Ieracitano et al., 2022; Jaiswal et al., 2019; Prakash et al., 2023a; Rehman et al., 2022). These models have hierarchical feature learning, robustness to variance, higher scalability, and capability due to the handling of millions of parameters to capture the image data with vast complexities. These models are trained with the centralized learning process, where the data processing and computational tasks are performed within the central server using the vast volume of the data. DL presents various efficient approaches for optimally analyzing the images, while it also poses significant challenges such as scalability issues, data privacy concerns, and high time complexity due to transferring the data to the centralized loca (...truncated)