WomenSafe: An Ultra-Lightweight Convolution–Vision Transformer Fusion Model for PCOS Diagnosis

Polycystic Ovary Syndrome (PCOS) affects approximately 8 to 13 percent of women of reproductive age worldwide. It is one of the leading causes of female infertility and is classified as a hormonal disorder that impacts women during their childbearing years. Common symptoms of PCOS include acne, irregular menstrual cycles, excessive body hair growth, and weight gain. Early diagnosis is crucial for effective symptom management and reducing associated health risks. However, the absence of channel-wise attention in deep learning models, particularly convolutional neural networks (CNNs), limits their ability to capture feature channel interdependencies efficiently. To enhance automatic PCOS identification, this study explores the integration of Squeeze-and-Excitation (SE) blocks with deep learning models and a Modified Vision Transformer (MViT). Using ultrasound images, we evaluate state-of-the-art models and propose a novel, lightweight Convolutional–Vision Transformer Fusion Model (WomenSafe). The WomenSafe model leverages global average pooling to identify the most relevant features by combining characteristics from MobileNetV2 with SE blocks and a MViT with SE blocks. As part of this study, we introduce WomenSafe, a proposed framework designed to improve PCOS classification accuracy using ultrasound imaging. The automated analysis of ultrasound images for early PCOS detection has yielded promising results, suggesting potential clinical applications. Experimental findings demonstrate that WomenSafe outperforms baseline models, achieving an accuracy of 95.31%, a precision of 93.06%, a recall of 94.37%, an F1-score of 93.71%, and a Matthews Correlation Coefficient (MCC) of 89.08%. These promising results underscore the effectiveness of WomenSafe in accurately classifying PCOS patients based on ultrasound images, making it a valuable diagnostic tool.