Blur Synthesis for Artifact Segmentation in Digital Pathology

Surya Achanta, Yashasvi Singh, NILANJAN CHATTOPADHYAY, Nitin Singhal

Keywords: Histopathology, Diffusion Models, Artifact Detection, Synthetic Data Generation, Image Segmentation

Abstract: Histopathology image datasets frequently suffer from artifacts like out-of-focus blur due to inconsistent microscopy procedures, significantly compromising the reliability of downstream analysis. While identifying and segmenting these artifacts is essential for robust quality control, the development of supervised detection models is hindered by a scarcity of pixel-level annotations for blurred regions. To address this data limitation, we introduce a novel framework for synthesizing realistic blur artifacts utilizing deep generative modeling. Unlike conventional statistical approaches, which fail to capture the stochastic and spatially variant nature of optical aberrations, our approach leverages Conditional Generative Adversarial Networks (cGANs) and Conditional Denoising Diffusion Probabilistic Models (cDDPMs). These models are trained to translate sharp histological images into their blurred counterparts, preserving the textural semantics of the tissue. We demonstrate that segmentation networks trained on this synthetically generated data exhibit superior generalization in identifying blur artifacts compared to models trained on statistically degraded data or limited data

Primary Subject Area: Image Synthesis

Secondary Subject Area: Segmentation

Registration Requirement: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 397