Accurate classification of cervical cytology images is essential for early detection of cervical cancer, particularly in low-resource clinical settings. While transfer learning has shown strong performance in medical image classification tasks [6, 8], the role of preprocessing techniques specific to cytological image characteristics remains underexplored. This study presents a comparative analysis of five medical image preprocessing techniques—ROI Extraction, stain normalization, denoising, contrast enhancement, and resampling—applied prior to transfer learning on a real-world dataset of 11,400 images. We evaluate three pretrained convolutional neural networks (ResNet50, InceptionV3, and DenseNet201), all fine-tuned across all layers to fully adapt to domain-specific features. Our experimental results show that ROI Extraction and resampling yield significant improvements in classification performance, with ResNet50 achieving the highest F1-score of 72.9 %. In contrast, general-purpose techniques such as denoising and stain normalization lead to a marginal or negative impact on performance. These findings highlight the importance of domain-specific preprocessing in enhancing the effectiveness of transfer learning for medical imaging tasks. The study provides practical insights for developing reliable and scalable cervical cancer screening systems in real-world environments.