摘要
在仅依赖组织病理图像且缺乏额外辅助检测的情况下,腹膜后软组织肉瘤小体积活检样本易导致观察者之间的判断差异,影响疾病亚型的整体诊断准确性。为了解决这一问题,从多中心收集了157张全切片图像(WSIs),涵盖去分化脂肪肉瘤、平滑肌肉瘤、恶性周围神经鞘瘤、未分化多形性肉瘤和高分化脂肪肉瘤五种疾病类别。基于上述WSIs,提出了基于单尺度图像与多尺度图像的两种模型集成方法,并利用ResNet18、EfficientNet B7和EfficientNet V2等深度学习模型进行训练。结果表明:两种模型集成方法均取得了较高的分类准确率,最佳模型在块级分析中达到82.27%的总体准确率,在全切片分析中达到80.95%。因此,所提方法能够有效辅助病理学家在临床实践中诊断腹膜后软组织肉瘤。
In the absence of additional auxiliary tests and relying solely on histopathological images,small-volume biopsy samples of retroperitoneal soft tissue tumors often lead to interobserver variability,impacting the overall diagnostic accuracy of disease subtypes.To address this issue,157 whole-slide images(WSIs)were collected from multiple centers,encompassing five disease categories:dedifferentiated liposarcoma(DDLP),leiomyosarcoma(LMS),malignant peripheral nerve sheath tumor(MPNST),undifferentiated pleomorphic sarcoma(UPS),and well-differentiated liposarcoma(WDLP).Based on these WSIs,two model ensemble methods were proposed:one based on single-scale images and the other on multi-scale images.Deep learning models,such as ResNet18,EfficientNet B7,and EfficientNet V2,were trained on the collected data.Results showed that both ensemble methods achieved high classification accuracy,with the best model achieving an overall accuracy of 82.27%in patch-level analysis and 80.95%in WSI-level analysis.Therefore,the proposed methods can effectively assist pathologists in the clinical diagnosis of retroperitoneal soft tissue tumors.
作者
姜永军
李红玲
阮萍
陈路
李艳春
胡庆
谢功勋
孟云鹤
JIANG Yongjun;LI Hongling;RUAN Ping;CHEN Lu;LI Yanchun;HU Qing;XIE Gongxun;MENG Yunhe(.School of Artificial Intelligence,Sun Yat-sen University,Zhuhai 519082,China;Foshan Traditional Chinese Medicine Hospital,Foshan 528000,China;Ruikang Hospital Affiliated to Guangxi University of Chinese Medicine,Nanning 530000,China;Yichang Central People’s Hospital,Yichang 443000,China;Hunan Provincial People’s Hospital,Changsha 410000,China)
出处
《中山大学学报(自然科学版)(中英文)》
北大核心
2025年第3期156-164,共9页
Acta Scientiarum Naturalium Universitatis Sunyatseni
基金
国家自然科学基金(61673390)。
关键词
腹膜后软组织肉瘤
深度学习
全切片图像
亚型诊断
retroperitoneal soft tissue sarcomas
deep learning
whole slide images
subtyping
