Ke Xinyi, Yang Moxuan, Chen Jingci, Hong Ruping, Wang Zheng, Wang Shuhao, Zhang Hui, Lu Junliang, Pan Boju, Gao Yike, Liu Xiaoding, Li Xiaoyu, Zhang Yang, Su Si, Wu Huanwen, Liang Zhiyong
Department of Pathology, Peking Union Medical College Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China.
Thorough Lab, Thorough Future, Beijing, China; Department of Physics, Capital Normal University, Beijing, China.
Mod Pathol. 2025 Jul;38(7):100764. doi: 10.1016/j.modpat.2025.100764. Epub 2025 Apr 6.
Accurate histopathological evaluation of pancreatic ductal adenocarcinoma (PDAC), including primary tumor lesions and lymph node metastases, is critical for prognostic evaluation and personalized therapeutic strategies. Distinct from other solid tumors, PDAC presents unique diagnostic challenges owing to its extensive desmoplasia, unclear tumor boundary, and difficulty in differentiating from chronic pancreatitis. These characteristics not only complicate pathological diagnosis but also hinder the acquisition of pixel-level annotations required for training computational pathology models. In this study, we present PANseg, a multiscale weakly supervised deep learning framework for PDAC segmentation, trained and tested on 368 whole-slide images (WSIs) from 208 patients across 2 independent centers. Using only image-level labels (2048 × 2048 pixels), PANseg achieved comparable performance with fully supervised baseline (FSB) across the internal test set 1 (17 patients/58 WSIs; PANseg area under the receiver operating characteristic curve [AUROC]: 0.969 vs FSB AUROC: 0.968), internal test set 2 (40 patients/44 WSIs; PANseg AUROC: 0.991 vs FSB AUROC: 0.980), and external test set (20 patients/20 WSIs; PANseg AUROC: 0.950 vs FSB AUROC: 0.958). Moreover, the model demonstrated considerable generalizability with previously unseen sample types, attaining AUROCs of 0.878 on fresh-frozen specimens (20 patients/20 WSIs) and 0.821 on biopsy sections (20 patients/20 WSIs). In lymph node metastasis detection, PANseg augmented the diagnostic accuracy of 6 pathologists from 0.888 to 0.961, while reducing the average diagnostic time by 32.6% (72.0 vs 48.5 minutes). This study demonstrates that our weakly supervised model can achieve expert-level segmentation performance and substantially reduce annotation burden. The clinical implementation of PANseg holds great potential in enhancing diagnostic precision and workflow efficiency in the routine histopathological assessment of PDAC.
胰腺导管腺癌（PDAC）的准确组织病理学评估，包括原发性肿瘤病变和淋巴结转移，对于预后评估和个性化治疗策略至关重要。与其他实体瘤不同，PDAC因其广泛的促纤维增生、肿瘤边界不清以及难以与慢性胰腺炎区分而呈现出独特的诊断挑战。这些特征不仅使病理诊断复杂化，还阻碍了训练计算病理学模型所需的像素级注释的获取。在本研究中，我们提出了PANseg，这是一种用于PDAC分割的多尺度弱监督深度学习框架，在来自2个独立中心的208例患者的368张全切片图像（WSI）上进行训练和测试。仅使用图像级标签（2048×2048像素），PANseg在内部测试集1（17例患者/58张WSI；PANseg在受试者工作特征曲线下面积[AUROC]：0.969 vs FSB AUROC：0.968）、内部测试集2（40例患者/44张WSI；PANseg AUROC：0.991 vs FSB AUROC：0.980）和外部测试集（20例患者/20张WSI；PANseg AUROC：0.950 vs FSB AUROC：0.958）上取得了与全监督基线（FSB）相当的性能。此外，该模型在以前未见过的样本类型上表现出相当大的通用性，在新鲜冷冻标本（20例患者/20张WSI）上的AUROC为0.878，在活检切片（20例患者/20张WSI）上的AUROC为0.821。在淋巴结转移检测中，PANseg将6名病理学家的诊断准确率从0.888提高到0.961，同时将平均诊断时间减少了32.6%（72.0分钟对48.5分钟）。这项研究表明，我们的弱监督模型可以实现专家级的分割性能，并大大减轻注释负担。PANseg的临床应用在提高PDAC常规组织病理学评估中的诊断精度和工作流程效率方面具有巨大潜力。
