Department of Pathology, University of Yamanashi, Chuo, Yamanashi, Japan.
Department of Pathology, University of Yamanashi, Chuo, Yamanashi, Japan.
Comput Biol Med. 2024 Nov;182:109149. doi: 10.1016/j.compbiomed.2024.109149. Epub 2024 Sep 18.
Sections stained in periodic acid-Schiff (PAS), periodic acid-methenamine silver (PAM), hematoxylin and eosin (H&E), and Masson's trichrome (MT) stain with minimal morphological discordance are helpful for pathological diagnosis in renal biopsy. Here, we propose an artificial intelligence-based re-stainer called PPHM-GAN (PAS, PAM, H&E, and MT-generative adversarial networks) with multi-stain to multi-stain transformation capability. We trained three GAN models on 512 × 512-pixel patches from 26 training cases. The model with the best transformation quality was selected for each pair of stain transformations by human evaluation. Frechet inception distances, peak signal-to-noise ratio, structural similarity index measure, contrast structural similarity, and newly introduced domain shift inception score were calculated as auxiliary quality metrics. We validated the diagnostic utility using 5120 × 5120 patches of ten validation cases for major glomerular and interstitial abnormalities. Transformed stains were sometimes superior to original stains for the recognition of crescent formation, mesangial hypercellularity, glomerular sclerosis, interstitial lesions, or arteriosclerosis. 23 of 24 glomeruli (95.83 %) from 9 additional validation cases transformed to PAM, PAS, or MT facilitated recognition of crescent formation. Stain transformations to PAM (p = 4.0E-11) and transformations from H&E (p = 4.8E-9) most improved crescent formation recognition. PPHM-GAN maximizes information from a given section by providing several stains in a virtual single-section view, and may change the staining and diagnostic strategy.
经过 PAS、PAM、H&E 和 MT 染色后，如果形态学上差异极小，则有助于肾脏活检的病理诊断。在这里，我们提出了一种基于人工智能的重染剂，称为 PPHM-GAN（PAS、PAM、H&E 和 MT 生成对抗网络），具有多染色到多染色的转换能力。我们在 26 个训练病例的 512×512 像素的小块上训练了三个 GAN 模型。通过人工评估，选择每个染色转换对中转换质量最佳的模型。Fréchet 初始距离、峰值信噪比、结构相似性指数测量、对比度结构相似性和新引入的域偏移初始分数被计算为辅助质量指标。我们使用 10 个验证病例的 5120×5120 个小块验证了诊断的实用性，用于主要肾小球和间质异常。转换后的染色有时优于原始染色，有助于识别新月体形成、系膜细胞增生、肾小球硬化、间质病变或动脉硬化。来自 9 个额外验证病例的 24 个肾小球中的 23 个（95.83%）可以转换为 PAM、PAS 或 MT，有助于识别新月体形成。PAM 染色的转换（p=4.0E-11）和从 H&E 的转换（p=4.8E-9）最能提高新月体形成的识别。PPHM-GAN 通过在虚拟单节视图中提供几种染色，最大化了给定切片的信息，并可能改变染色和诊断策略。
