A Cervical Lesion Recognition Method Based on ShuffleNetV2-CA.

Abstract

Cervical cancer is the second most common cancer among women globally. Colposcopy plays a vital role in assessing cervical intraepithelial neoplasia (CIN) and screening for cervical cancer. However, existing colposcopy methods mainly rely on physician experience, leading to misdiagnosis and limited medical resources. This study proposes a cervical lesion recognition method based on ShuffleNetV2-CA. A dataset of 6,996 cervical images was created from Hebei University Affiliated Hospital, including normal, cervical cancer, low-grade squamous intraepithelial lesions (LSIL, CIN 1), high-grade squamous intraepithelial lesions (HSIL, CIN 2/CIN 3), and cervical tumor data. Images were preprocessed using data augmentation, and the dataset was divided into training and validation sets at a 9:1 ratio during the training phase. This study introduces a coordinate attention mechanism (CA) to the original ShuffleNetV2 model, enabling the model to focus on larger areas during the image feature extraction process. Experimental results show that compared to other classic networks, the ShuffleNetV2-CA network achieves higher recognition accuracy with smaller model parameters and computation, making it suitable for resource-limited embedded devices such as mobile terminals and offering high clinical applicability.