AI-Assisted PCOS Detection from Ultrasound Scans: Benchmarking VGG16, ResNet50, and EfficientNet

Abstract

Polycystic Ovary Syndrome (PCOS) is one of the most common disorders of the endocrine system in women of reproductive age, and one of the most common means to diagnose the conditions is the use of transvaginal ultrasounds. Still, the images can be subjective, and their interpretation can greatly depend on the skill level of the interpreter and the experience of the examiner. This study seeks to improve consistency and accuracy in diagnosis with the use of deep learning, designed to be more accurate computer-aided systems, on a sample of 11,784 practical ultrasound images. For this study, three deep learning models using different architectures of convolutional neural networks, called VGG16, ResNet50, and EfficientNetB0, were compared in the classification of ultrasound images and the respective PCOS features. VGG16 was the best model with respects to accuracy, reaching 95.40% and an overall F1-measure of 95.54%. Moreover, EfficientNetB0 was the most capable model in delineating the core patterns of PCOS with an AUC of 96.97%. Strong AUC values were a sign of strong ability to distinguish patterns. Each model in the study was evaluated using a confusion matrix, an ROC Curve, and validation loss metrics to assess reliability. Even with the common challenges of noisy images and grayscale resolution inconsistencies, results of the study demonstrate the feasibility and success of the implementation of deep learning for PCOS diagnosis.