A Lightweight Method for Cervical Cancer Classification Using Preprocessing Pipeline and Attention-Guided Shallow-CNN

Abstract

Cervical cancer continues to be one of the most common cancers in women globally, and early detection by means of automated image classification can lead to significant improvement in survival rates. This paper introduces a novel and effective deep learning-based technique for automatic cervical cell image classification into six diagnostic classes. The model starts with a strong image preprocessing pipeline, including Contrast-Limited Adaptive Histogram Equalization (CLAHE) to enhance the visibility of lesion areas, then uniform resizing and z-score normalization to normalize image input features. These preprocessing methods ensure the input of more enhanced and normalized image data to the model, reducing the impact of contrast and scale variation. Given that the classes in the dataset are unbalanced, targeted data augmentation is used for balancing classes before training. The transformation techniques of rotation, flipping, zooming, and brightness adjustments are applied to generate synthetic samples for classes that are underrepresented, thereby reducing overfitting and improving the generalization of the model.
A shallow Compact Convolutional Neural Network (CNN) is engineered to learn discriminative features without a high computational cost as required for implementation in low-resource settings. For added classification accuracy, a new Attention-Guided Feature Modulation (AGFM) block is added to the network. This process learns to modulate spatial and channel-wise salient features and inhibit unnecessary background noise, thereby channelling the model's attention towards dysplastic areas in cervical images. Experimental results show that the designed framework is highly accurate and generalizable for multi-class cervical cancer classification. Its light-weight nature, coupled with attention-guided improvements, positions it as an ideal candidate for real-time diagnostic assistant systems, particularly in low-resource healthcare environments.