Zheng Xueyi, Zheng Ke, Wen Yongqin, Meng Jiajia, Zhang Xinke, Wen Xiaobo, Zhao Zihan, Zheng Chengyou, Cai Xiaoxia, Lin Jiliang, Chen Jiewei, Duan Jinling, Jiang Liwen, Yuan Wei, Li Xiaomei, Xie Dan, Cai Yubo, Zhang Jiangyu, Cai Muyan
Department of Pathology, State Key Laboratory of Oncology in South China, Guangdong Provincial Clinical Research Center for Cancer, Sun Yat-sen University Cancer Center, Guangzhou, China.
Department of Pathology, The Tenth Affiliated Hospital, Southern Medical University, Dongguan People's Hospital, Dongguan, China.
NPJ Digit Med. 2025 Jul 20;8(1):460. doi: 10.1038/s41746-025-01808-7.
Intraoperative frozen section diagnosis provides essential, real-time histological insights to guide surgical decisions. However, the quality of these time-sensitive sections is often suboptimal, posing significant diagnostic challenges for pathologists. To address these limitations, we utilized over 6700 whole slide images to develop GAS, a comprehensive platform comprising three modules: Generation, Assessment, and Support modules. The Generation module, based on a GAN-driven multimodal network guided by FFPE-style text descriptions, demonstrated effective enhancement of frozen section quality across various organs. The Assessment module, which fine-tuned quality control models using pathological foundation models, showed substantial improvements in microstructural quality for the generated images. Validated through a prospective study (ChiCTR2300076555) on the human-AI collaboration software, the Support module demonstrated that GAS significantly boosted diagnostic confidence for pathologists. In summary, this study highlights the clinical utility of the GAS platform in intraoperative diagnosis and establishes a new paradigm for integrating end-to-end AI solutions into clinical workflows.
术中冰冻切片诊断提供了重要的实时组织学见解，以指导手术决策。然而，这些对时间敏感的切片质量往往不尽人意，给病理学家带来了重大的诊断挑战。为了解决这些局限性，我们利用了6700多张全切片图像来开发GAS，这是一个由三个模块组成的综合平台：生成模块、评估模块和支持模块。基于由FFPE风格文本描述引导的GAN驱动的多模态网络的生成模块，在各个器官中均有效提高了冰冻切片质量。使用病理基础模型对质量控制模型进行微调的评估模块，在生成图像的微观结构质量方面有了显著改善。通过对人机协作软件的前瞻性研究（ChiCTR2300076555）进行验证，支持模块表明GAS显著提高了病理学家的诊断信心。总之，本研究突出了GAS平台在术中诊断中的临床实用性，并建立了将端到端人工智能解决方案整合到临床工作流程中的新范例。
