A federated learning based efficient approach to detect cervical cancer using PAP-SMEAR images

Abstract

Cervical cancer is one of the most common malignancies and easily preventable through early diagnosis of pap smear images which results in a lower death rate among women throughout the world.The typical examination of pap smear images is tiring and has a lot of human mistakes. Additionally, having less privacy due to data centralization is a significant concern.In the present thesis work, proposes an efficient privacy-preserving federated learning based framework for cervical cancer detection using Pap smear images.Detection of cervical cancer is first done with comparative experimental analysis for pretrained and hybrid lightweight models such as VGG16, ResNet18, ResNet50, DenseNet121, EfficientNetB0, Vision Transformer and MobileNetV2 in iid and non-iid data distribution.In addition, FedPapVisionNet, a lightweight mechanism for federated modelling was introduced which improves classification performance while safeguarding privacy.It merges the efficient convolutional operation and attention mechanism.With only 3,90,215 trainable parameters, it achieves impressive diagnostic accuracy on datasets.Ultimately, the use of federated optimization techniques such as FedAvg, FedProx, and Scaffold allows for the assessment of convergence and robustness in homogeneous and heterogeneous data.Experimental results demonstrate that FedPapVisionNet achieved accuracy of 95.68% in IID distribution, 86% using FedProx algorithm and 90% using SCAFFOLD strategy in non-IID distribution, highlighting its effectiveness for accurate, scalable and privacy-aware cervical cancer detection in real-world healthcare applications.