多肿瘤淋巴管癌栓自动检测模型效果评价研究

Evaluation of the effect of automatic detection model for lymphovascular infiltration inmultiple tumors

目的:测试并分析全局库(GLB)的淋巴管癌栓(lymphovascular infiltration,LVI)目标检测模型的检测效果及泛化能力。方法:随机选取西安交通大学第一附属医院病理科多癌种的淋巴管癌栓阳性及阴性病例共131例,所有切片均采用HE以及D2-40免疫组化染色。实验分组包括病理医师人工阅片组、算法组以及算法辅助病理医师三组。分别统计三组检测LVI结果以及检测时间。结果:比较病理医师、纯算法和算法辅助病理医师的一致性显示:131例病例中显示三种模式下一致性Kappa值分别为较好(0.709、0.786、0.796)、轻微(0.208)和中等一致(0.412)。病理医师出错的最常见的原因是漏诊(n=8),第二是由于IHC染色问题(包括6例染色不足,而3例未着色)。其余包括2例组织裂隙,1例小静脉(D2-40未着色),3例原因不明。纯算法判断错误的原因,一方面是由于病理切片或染色技术引起的,例如游离小组织(15/63),非特异性或染色不足(14/63),气泡(8/63)等。另一方面是由于某些LVI的组织形态和检测目标的相似性而导致算法误判,例如腺体(14/63),血管(13/63)和周围小神经(5/63)等。三种模式下所花费时间分别为59.28 s、82.40 s和10.43 s。结论:我们的结果虽然证实了算法辅助病理诊断提高诊断效率的能力,然而,考虑到算法辅助诊断的落地,病理切片的质量控制问题不容忽视,此外应考虑是否按照不同系统或疾病开展多层次算法设计,并且尽可能通过大数据的积累以覆盖多种检测目标类似物的必要性。

Objective:To test and analyze the detection effect and generalization ability of the global library(GLB)target detection model of lymphovascular infiltration(LVI).Methods:131 cases of LVI positive and negative from the Department of Pathology of the First Affiliated Hospital of Xi'an Jiaotong University were randomly selected.All slides were stained with HE and D2-40 immunohistochemistry.The experimental group includes three groups:pathologists'manual reading group,algorithm group and algorithm assisted pathologists.The LVI results and detection time of the three groups were counted respectively.Results:The comparison of the consistency among pathologists,pure algorithm and algorithm assisted pathologists showed that in 131 cases,the consistency Kappa values under the three modes were good(0.709,0.786,0.796),slight(0.208)and moderate(0.412),respectively.The most common reason for pathologists'errors is missed diagnosis(n=8),and the second is due to IHC staining problems(including 6 cases of insufficient staining,while 3 cases of significant non staining).The rest included 2 tissue fissures,1 small vein(D2-40 not stained),and 3 unknown causes.On the one hand,the reason for the error of pure algorithm judgment is due to pathological section or staining technology,such as free small tissues(15/63),non-specific or insufficient staining(14/63),bubbles(8/63),etc.On the other hand,the similarity between the tissue morphology of some LVIs and the detection targets leads to algorithm misjudgment,such as glands(14/63),blood vessels(13/63),and peripheral small nerves(5/63).The time spent in the three modes is 59.28 s,82.40 s and 10.43 s respectively.Conclusion:Although our results confirm the ability of algorithm assisted pathological diagnosis to improve diagnostic efficiency,the implementation of algorithm assisted diagnosis should fully consider the quality control of pathological slides and the necessity of developing multi-level design,and try to accumulate large data to cover multiple target analogues.