Comparing Image Restoration Methods for Cytology

Cervical cytology screening depends on image quality. Blur and noise in liquid-based cytology (LBC) preparations can obscure cellular details, leading to diagnostic uncertainty or missed abnormalities. Computational image restoration offers a path to improving image quality without repeat sampling.

Total Variation (TV) regularization is a classical optimization-based method that reduces noise while preserving edges. It is computationally efficient and requires no training data, making it practical for clinical deployment.

UNet architectures represent the deep learning approach — trained on pairs of degraded and clean images, they learn to restore quality end-to-end. Performance depends heavily on the quantity and quality of training data.

DiffPIR applies diffusion-based reconstruction, treating restoration as a reverse diffusion process guided by the degraded image. It produces visually impressive results but requires significantly more compute.

UNet achieved the best quantitative metrics (PSNR and SSIM), while DiffPIR produced results preferred by pathologists in qualitative review. TV regularization, while lower in absolute performance, offers the advantage of predictable behavior without training data dependencies.