Akademik Veri Yönetim Sistemi
9th International Conference on Computer Science and Engineering, UBMK 2024, Antalya, Türkiye, 26 - 28 Ekim 2024, ss.371-376, (Tam Metin Bildiri)
Mitosis on H&E-stained (Hematoxylin and Eosin) images is an important prognostic marker for evaluating the tumor's aggressiveness and providing comprehensive and reliable information for accurate diagnosis and treatment. The variability in dosage rate and different fading rates in H&E stains lead to challenges in histopathological analysis. Standardizing the staining process and understanding the factors that influence stain stability is critical for consistency and accuracy in diagnosis. This paper proposes an improved stain normalization method with DDPM-based (Denoising Diffusion Probabilistic Models) diffusion model to detect mitosis with high accuracy in histopathological images. One of the challenges in the DDPM-based diffusion model is incorrect transfer, where the model may confuse hematoxylin with eosin due to the high diversity nature of DDPM-based diffusion models. A stain separation method using CD-SNMF (Color Deconvolution-Sparse Non- Negative Matrix Factorization) is proposed to address this issue. This method normalizes the mitosis in images in CCMCT (Canine Mast Cell Tumors) and MIDOG21 (Mitosis Domain Generalization Challenge 2021) datasets by this stain normalization pathway with decent performance.