Attention-Enhanced Deep Convolutional Denoising Autoencoder for Cervical Cancer Image Quality Improvement

Abstract

Cervical cancer is the main cause of death among women worldwide, and catching it early can make all the difference. Pap-smear images are a cornerstone of screening, but in practice they’re often affected by noise tiny specks, uneven lighting or blur that can throw off both human experts and automated algorithms. To address this, we’ve built an Attention-Enhanced Deep Convolutional Denoising Autoencoder (AE-DCDA). By weaving an attention mechanism into a classic encoder decoder structure, our model learns to focus on the important cell structures and suppress the noise around them, preserving the fine details that matter for diagnosis. We tested AE-DCDA on the Herlev dataset, which originally includes 917 cervical cell images spanning seven classes. To give the model more varied samples, we applied a range of image augmentation techniques such as rotations, flips and small shifts, effectively enlarging our training pool. When we evaluate noisy images held back, our denoiser pushed the Peak Signal to Noise Ratio up to 35.2 dB and drove the Mean Squared Error down to 0.0003, notable gains over conventional filters. In practice, that means clearer cell boundaries, crisper nuclei and fewer artifacts, paving the way for more reliable downstream segmentation or classification.