Fine-Tuning Strategies for Lightweight Models in Cervical Cells Classification

Abstract

Cervical cancer remains one of the most common and preventable health threats to women worldwide. Early diagnosis significantly improves treatment outcomes. Automated abnormal cell screening facilitates early detection and diagnosis of cervical cancer, aiding in disease prevention and improving patient survival rates. Prior research on cervical cell imaging has primarily focused on developing and training deep learning models tailored specifically for cervical cancer classification. However, a more computationally efficient approach that also reduces the risk of overfitting is transfer learning through fine-tuning a lightweight pre-trained model. This study proposes various fine-tuning strategies applied to cervical cancer cytology datasets, specifically CRIC and LBC, to identify the most optimal approach. Experimental results demonstrate that a hybrid fine-tuning strategy achieves a classification accuracy of up to 99%. In addition to offering insights into the application of transfer learning on lightweight models for cervical cell classification, this study establishes a strong baseline for future research in automated cervical cancer diagnosis.