EffiDenseGenOp: Ensemble Transfer Learning With Hyperparameter Tuning Using Genetic Algorithm Optimization for PCOS Detection From Ultrasound Sonography Images

• Published in: IEEE access Vol. 13; pp. 54285 - 54312
• Main Authors: Ghosh, Ananya, Srinivasan, Kathiravan
• Format: Journal Article
• Language: English
• Published: IEEE 2025
• Subjects: Accuracy; Analytical models; Biomedical imaging; Deep learning; genetic algorithm optimization; Genetic algorithms; hyperparameter tuning; Medical diagnostic imaging; Polycystic ovary syndrome; Predictive models; Transfer learning; Tuning; Ultrasonic imaging
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2025.3553895

In this research, we present the revolutionary 'EffiDenseGenOp' framework for Polycystic Ovary Syndrome (PCOS) detection, leveraging the amalgamation of Ensembled Transfer Learning Models. By harnessing the synergies of Ensembled EfficientNetB7 and DenseNet201, our approach transcends conventional models, offering a robust solution for PCOS detection. Notably, we introduce a Genetic Algorithm-based hyperparameter tuning mechanism, optimizing the model configuration to ensure superior generalization and performance. Our contributions encompass a meticulous comparative analysis, pitting machine learning models, deep learning models, transfer learning models, and our proposed Ensembled Transfer Learning Models against each other. The ensemble technique strategically captures complementary patterns and features from each base model, significantly amplifying the overall predictive power. Moreover, we conduct a comprehensive exploration of hyperparameters, employing extensive tuning to enhance model performance and generalization. The efficiency of Genetic Algorithm is underscored through a rigorous comparative analysis. Additionally, a novel Fuzzy Inference System is introduced for image quality enhancement, designed after meticulous examinations of image behavior under varying noise levels and membership functions. Our model undergoes rigorous training through diverse image variations, employing data augmentation techniques and noise addition. Performance evaluation reveals superior accuracy (99.58%), precision (98.87%), recall (99.20%), F1-score (99.01%), and AUC-ROC score (98.97%), substantiated by detailed analyses including confusion matrices and AUC-ROC curves. Compared to existing models, our proposed model outperforms several state-of-the-art techniques, such as VGGNet16, PCOSNet, and ResNet50, with an accuracy of 99.95%, highlighting a significant improvement in PCOS detection performance. Robustness is ensured through exhaustive K-fold cross-validation, while visualization techniques like Class Activation Maps shed light on the model's interpretability. The research attains highly efficient PCOS detection, achieving elevated accuracy levels validated through a detailed ablation study.