基于多种染色病理图像的非炎性主动脉中膜变性计算机辅助诊断方法

  目的  探讨建立非炎性主动脉中膜变性(medial degeneration, MD)患者多种染色病理图像计算机辅助诊断模型的可行性。  方法  回顾性收集2018年7—12月首都医科大学附属北京安贞医院诊治胸主动脉瘤及夹层患者非炎性病变的主动脉手术标本病理切片。将其以400倍率扫描为全视野数字图像(whole slide image, WSI)后由2名病理科医师对病变进行标注。按6:1的比例将标注后的WSI图像随机分为训练集和测试集。采用训练集数据对SE-EmbraceNet进行训练，构建多种染色病理图像MD多分类包括层内型黏液样细胞外基质聚集(intralamellar mucoid extracellular matrix accumulation，MEMA-I)、穿透型黏液样细胞外基质聚集(translamellar mucoid extracellular matrix accumulation，MEMA-T)、弹力纤维断裂和/或缺失(elastic fiber fragmentation and/or loss，EFFL)和平滑肌细胞核缺失(smooth muscle cell nuclei loss，SMCNL)]模型，并采用准确率、灵敏度、精确率、F1值对模型的分类效果进行评价。  结果  共入选符合纳入和排除标准的胸主动脉瘤及夹层患者非炎性主动脉病变手术标本病理切片530张。总提取5265组图像，每组均包含同一病变部位HE染色、特殊染色(弹力纤维/VanGieson、Masson、阿辛蓝/过碘酸雪夫)及平滑肌肌动蛋白5种染色病理图像。其中，训练集图像4513组，包括SMCNL 987组、EFFL 2013组、MEMA-I 1337组及MEMA-T 176组；测试集图像752组，包括SMCNL 166组、EFFL 335组、MEMA-I 222组及MEMA-T 29组。模型在测试集中显示出良好的MD分类性能，整体准确率为96.54%(726/752)。其中，对EFFL的识别能力最强，准确率、灵敏度、精确率和F1值均≥98.51%，其次为SMCNL，各评价指标均≥97.59%，亦具有较好的分类能力。  结论  本研究构建的基于多种染色病理图像MD分类模型具有较高的分类准确性和较好的泛化能力，有望应用于非炎性主动脉病变的辅助诊断。

Computer-aided Diagnostic Methods for Medial Degeneration in Non-inflammatory Aorta Based on Multi-stained Pathological Images

  Objective  To explore the feasibility of establishing a computer-aided diagnostic model of multi-stained pathological images in patients with non-inflammatory aortic medial degeneration(MD).  Methods  In this study, pathological sections of aortic surgical specimens for non-inflammatory lesions from patients with thoracic aortic aneurysms and dissections were retrospectively collected at the Beijing Anzhen Hospital, Capital Medical University from July to December 2018. The lesions were scanned under ×400 magnification as whole slide images(WSI) and then annotated by two pathologists. The annotated WSI images were randomly split into training and test sets in a 6:1 ratio. SE-EmbraceNet was used to train the data to construct a multi-classification model for MD of multi-stained pathology images, including intralamellar mucoid extracellular matrix accumulation (MEMA-I), translamellar mucoid extracellular matrix accumulation (MEMA-T), elastic fiber fragmentation and/or loss(EFFL) and smooth muscle cell nuclei loss (SMCNL). The classification effect of the model was evaluated based on the test set data, and the results were expressed in terms of accuracy, sensitivity, precision, and the F1 value.  Results  Totally 530 pathological slides of non-inflammatory aortic lesion surgical specimens from patients with aortic aneurysm and dissection were included. Extracted 5265 sets of images, each containing 5 stained pathological images of the same lesion site: HE staining, special staining (elastic fiber/VanGieson, Masson, Alcian blue/periodic acid Schiff) and smooth muscle actin staining. There were 4513 sets of training images, including 987 SMCNL, 2013 EFFL, 1337 MEMA-I, and 176 MEMA-T; and 752 test images including 166 SMCNL, 335 EFFL, 222 MEMA-I, and 29 MEMA-T. The overall performance of the model in the test set showed good results, with an accuracy of 96.54%(726/752). The model had the best classification performance for EFFL, with accuracy, sensitivity, precision, and F1 value all ≥98.51%. The model also had a great classification ability for SMCNL, with all evaluated indexes≥97.59%.  Conclusion  The multi-stained pathology image-based MD classification model constructed in this study has high classification accuracy and good generalization ability, which has the potential to be applied to assist in the diagnosis of the non-inflammatory aortic lesion.