Comparative analysis of cervical cancer classification of DPAGCHE-enhanced Pap smear images using convolutional neural network models

Cervical cancer remains a significant cause of female mortality worldwide, primarily due to abnormal cell growth in the cervix. This study proposes an automated classification method to enhance detection accuracy and efficiency, addressing contrast and noise issues in traditional diagnostic approaches. The impact of image enhancement on classification performance is evaluated by comparing transfer learning-based Convolutional Neural Network (CNN) models trained on both original and enhanced images. This study employs transfer learning with pre-trained CNNs to classify preprocessed Pap smear images into three categories. Data augmentation, including rotations, flips, and shifts, enhances variability and prevents overfitting. The OneCycle learning rate schedule dynamically adjusts the learning rate, improving training efficiency. To enhance image quality, the Denoised Pairing Adaptive Gamma with Clipping Histogram Equalization (DPAGCHE) method improves contrast and reduces noise. The evaluation involves five pre-trained CNN models and the publicly available Herlev dataset, implemented in MATLAB Online. The ResNet50 model trained on the DPAGCHE-enhanced dataset achieves the highest classification performance, with 84.15% accuracy, along with improved specificity, recall, precision, and F1-score. ResNet50’s residual connections mitigate vanishing gradient issues and enhance deep feature extraction. Accordingly, the DPAGCHE preprocessing significantly improves classification performance, leading to a 53.65% increase in F1-score and 44.29% in precision. In contrast, the Baseline CNN reaches only 66.67% accuracy, highlighting the advantage of deeper architectures combined with enhanced preprocessing. These findings suggest integrating DPAGCHE-enhanced preprocessing with deep learning improves automated cervical cancer detection. In particular, ResNet50 demonstrates the best performance, reinforcing the effectiveness of contrast enhancement and noise reduction in aiding classification models.