基于深度迁移学习的卵巢上皮性癌术中冰冻切片病理诊断模型研究

Study on pathological diagnosis model of intraoperative frozen sections of epithelial ovarian cancer based on deep transfer learning

目的:探索基于深度迁移学习的卵巢上皮性癌术中冰冻切片病理诊断模型,并评估其应用价值。方法:收集2021年1月-2022年12月我院病理科卵巢上皮性癌术中冰冻切片25张以及术后石蜡切片121张,共5种组织学类型(浆液性癌、黏液性癌、子宫内膜样癌、透明细胞癌及转移性印戒细胞癌),扫描成数字病理全场图片。从切片上切出图像块数据集,标注框外的图像块归为Other类,共六类图像块数据集,按3∶1∶1划分为训练集、测试集和验证集。使用冰冻切片数据集训练从零学习模型,使用石蜡切片数据集训练预训练模型。对预训练模型在冰冻切片数据集上进行迁移学习优化得到深度迁移学习模型。对比使用和不使用迁移学习模型两种条件下,训练的六分类数据集在对应测试集中的表现。结果:迁移学习模型的各项指标相较于从零学习模型均有明显提升,迁移学习模型对于六类滑块的预测准确率均达到90%。结论:迁移学习模型在卵巢上皮性癌术中冰冻切片病理图像分类模型中具有较高的准确率和稳定性,具备较好的推广性。

Objective:To explore the pathological diagnostic model for intraoperative frozen sections of epithelial ovarian cancer using deep transfer learning and to assess its practical value.Methods:From January 2021 to December 2022,25 intraoperative frozen sections and 121 postoperative paraffin sections of ovarian epithelial cancer from department of pathology of our hospital were collected.These included five histopathological types(serous carcinoma,mucinous carcinoma,endometrioid carcinoma,clear cell carcinoma,and metastatic signet ring cell carcinoma),which were scanned into whole-slide digital pathology images.Image block datasets were extracted from these sections,with those outside the annotation frames categorized as"Other",resulting in six distinct image block datasets.The datasets were split into training,testing,and validation sets in a 3∶1∶1 ratio respectively.A model from scratch was trained using the frozen section dataset,while a pretrained model was trained using the paraffin section dataset.The pretrained model was further optimized on the frozen section dataset through transfer learning to develop a deep transfer learning model.The performance of the trained six-category datasets was compared in corresponding test sets with and without the application of the transfer learning model.Results:Compared to the model trained from scratch,the transfer learning model presented a significant improvement across all metrics,achieving a prediction accuracy of 90%for all six classes of image blocks.Conclusion:The transfer learning model exhibits a high level of accuracy and stability in the classification of pathological images from intraoperative frozen sections of epithelial ovarian cancer,indicating good potential for wider implementation.