Author ORCID Identifier

https://orcid.org/0009-0003-5570-2368

Date of Award

17-8-2025

Document Type

Thesis

School

School of Computing

Programme

Ph.D.-Doctoral of Philosophy

First Advisor

Dr.T.Suriya Praba

Keywords

Poly-Cystic Ovary Syndrome, Machine Learning, Deep Learning, Federated Learning, Differential Privacy

Abstract

A hormonal disorder, Poly-Cystic Ovary Syndrome (PCOS) usually affects women during the reproductive age. It is characterised by imbalances in hormones, particularly a rise in the female body's androgen level (male hormone) and enlarged ovaries with small cysts. PCOS can cause ovarian cysts, weight gain, acne, excessive hair growth, insulin resistance, and irregular menstrual cycles along with other health problems. While the exact origin of PCOS is uncertain and its symptoms are unclear, diagnosing PCOS in real-world conditions is a difficult task. Therefore, prompt and precise PCOS diagnosis is essential for efficient treatment and for averting long-term issues.

Clinicians typically use clinical, hormonal and ultrasound ovary images to manually analyse PCOS, although this method is labor-intensive and unreliable. Thus, the development of effective and automated PCOS classification models becomes essential for optimizing time and medical resources. This work uses the predictive power of Machine Learning (ML) and Deep Learning (DL) approaches to address the need for enhanced PCOS categorization. The goal is to investigate, create and evaluate ML and DL models that can accurately categorize PCOS from clinical, hormonal and ultrasound ovarian images, ultimately improving diagnostic accuracy and enabling timely intervention.

In machine learning, protecting the privacy of sensitive information has always been a top concern, particularly in the healthcare domain. Data may be exposed during various stages of model implementation including data collection, training and even after the release of a trained model. Thus, it is crucial to prevent data leakage and ensure patient privacy concerning Personally Identifiable Information (PII), ML model updates and Personal Health Information (PHI). To address these issues, a Fog-based Federated Learning approach is adopted, enabling collaborative learning where only gradients or updates from locally trained models are shared with the global server.

This research initially proposes a hybrid machine learning model using clinical and hormonal datasets for PCOS classification. Ensemble Feature Selection (FS) methods are used to identify the most significant indicators by selecting relevant features from a large feature set. A common issue in real-world applications is the instability of FS algorithms when applied repeatedly on the same or slightly modified datasets. Therefore, assessing FS robustness is crucial. In this study, Jensen–Shannon Divergence (JSD), an information-theoretic measure, is used along with ensemble FS to manage diverse outputs such as complete rankings, top-k lists and partial rankings. The resulting high-stability features are then used for training, and the proposed hybrid model achieved a remarkable accuracy of 97.81% using the AdaBoost classifier.

Although hormonal data contributes to PCOS diagnosis, it alone is insufficient. Ultrasound ovary images provide crucial visual information such as follicle size, count, shape and texture, essential for accurate diagnosis. A CNN-based Automated High-Precision PCOS Detection model is developed using ESRGAN for image resolution enhancement and SAM for cyst segmentation. Using VGG-19 with enhanced and segmented images, the model achieved an impressive 99.31% accuracy. Furthermore, a Differential Privacy (DP)-enabled Federated Learning framework is implemented for decentralized model training to ensure privacy and prevent data leakage. To balance the trade-off between privacy budget and utility, DP is applied only to the top-k participants, enabling the global model to achieve 87.29% accuracy with 0.3 top participants. The proposed model is also validated against two major attacks: gradient-based data reconstruction and model inversion.

Download

Share

COinS