基于多层病理学监督的乳腺癌H&E生成虚拟Ki-67图像研究OA

Regarding the problem of time-consuming of Ki-67 immunohistochemical staining and weak pairing between H&E and Ki-67 images in breast cancer,we proposed a multi-level pathology-guided supervised generative adversarial network(MPAS-GAN)to generate high-quality virtual Ki-67 images from H&E counterparts to assess the expression distribution of the Ki-67 biomarker.Firstly,the multi-level pathology-guided supervision framework was introduced in MPAS-GAN to solve the tissue misalignment at the macro-feature level,through the confidence-weighted optimal transport alignment.Finally,the diagnostic information at the key patho-logical semantic level was restored through the consistency constraint of Ki-67 pathological information,and the nuclear morphology at the basic cell structure level was retained through the consistency constraint of pathological cell structure.Experimental results on the public MIST and IHC4BC datasets demonstrated that MPAS-GAN significantly outperformed existing state-of-the-art methods across structural similarity index measure(SSIM),peak signal-to-noise ratio(PSNR),Fréchet inception distance(FID),learned percep-tual image patch similarity(LPIPS)metrics.Furthermore,it achieved the highest consistency in the quantitative correlation analysis of Ki-67 positive regions.This research can generate visually realistic and pathologically reliable virtual Ki-67 images,which can effec-tively solve the problem of weakly paired medical image translation,and is expected to provide a more efficient and reliable tool for the digital pathological diagnosis of breast cancer.