目的 擬通過深度學(xué)習(xí)技術(shù)，建立小鼠胃鱗狀細(xì)胞癌輔助診斷模型，以提高病理診斷的準(zhǔn)確性和一致性。方法 收集致癌性研究中小鼠胃鱗狀細(xì)胞癌組織93例和正常小鼠胃組織56例，掃描成數(shù)字切片后，進(jìn)行半自動(dòng)化數(shù)據(jù)標(biāo)注。對(duì)所有數(shù)據(jù)進(jìn)行組織提取、偽影去除以及良性上皮區(qū)域剔除等預(yù)處理后，按照8∶1∶1的比例隨機(jī)分為訓(xùn)練集、驗(yàn)證集和測(cè)試集?；贖ALO AI平臺(tái)構(gòu)建DenseNet算法模型用以識(shí)別胃鱗狀細(xì)胞癌區(qū)域和非鱗狀細(xì)胞癌區(qū)域。采用精確率（Pr）、召回率（Re）及F1-Score對(duì)構(gòu)建的算法模型進(jìn)行性能評(píng)估。結(jié)果 構(gòu)建的DenseNet算法模型在測(cè)試集中的總體Pr為0.904，召回率為0.929，F(xiàn)1-Score為0.916。結(jié)論 建立的DenseNet算法模型對(duì)于輔助診斷小鼠胃鱗狀細(xì)胞癌具有良好的應(yīng)用前景。

Objective To establish a assisted diagnosis model for mouse gastric squamous cell carcinoma, by implementing deep learning technology to improve the accuracy and consistency of pathological diagnosis. Methods A total of 93 cases of gastric squamous cell carcinoma tissue and 56 cases of normal mouse gastric tissue were collected form a carcinogenicity study. After scanning into digital slide images, semi-automated data annotation was performed. After preprocessing all data with tissues detection, artifact removal, and benign epithelial region removal, they were randomly divided into training set, validation set, and test set at a ratio of 8∶1∶1. Construct a DenseNet algorithm model based on the HALO AI platform to identify areas of gastric squamous cell carcinoma and non-squamous cell carcinoma. Evaluate the performance of the constructed algorithm model using precision, recall, and F1-score. Results The overall accuracy, recall and F1 score of the DenseNet algorithm model in the test set were 0.904, 0.929 and 0.916, respectively. Conclusion The DenseNet algorithm model established in this study has good application prospects for assisting diagnosis of gastric squamous cell carcinoma in mouse.

R965.1

中檢院學(xué)科帶頭人課題（2021X2）;藥品監(jiān)管科學(xué)全國(guó)重點(diǎn)實(shí)驗(yàn)室課題（2023SKLDRS0127）