生物组织偏振成像系统误差校正及甲状腺病理识别

打开文本图片集

关键词：偏振成像；误差校正；Mueller矩阵矢量参数；纹理特征；随机森林中图分类号：0436.3 文献标识码：Adoi：10.37188/OPE.20253320.3163 CSTR：32169.14.OPE.20253320.3163

Abstract： In order to improve the accuracy and stability of the system in the detection of pathological tissue samples，and to explore its application potential in the auxiliary diagnosis of thyroid cancer，a multi-factor cross-module error correction model was proposed to improve the accuracy and stability of the system in the detection of pathological tissue samples. Firstly，the main sources of system errors are analyzed，the error transfer optical path model is established by analytical method and numerical reconstruction method， and a multi-factor cross-module error correction model with 16 calibration parameters is constructed.Secondly，the nonlinear least squares fiting method is used to calibrate 16 parameters. According to the error correction model，the Mueler matrix of the airand blank slices is detected to evaluate the detection accuracy.Then，using the unstained sections of papillary thyroid carcinoma and nodular goiter as samples，four vector parameters （Δ，P，D，R） were extracted by Mueller matrix polarization decomposition method， and the texture features of each vector parameter image were extracted，and two classification models of random forest and support vector machine were constructed to obtain confusion matrix and ROC curve. Finally，the classification effect was evaluated bycalculating Precision，Recall，F1-score，and AUC.The experimental results show that the calibration accuracy is increased by 12% ，the calibration stability is increased by 21.5% ，and the detection accuracy is increased by 59% . The classification effect of random forest was better than that of support vector machine，and the classification effect of parameter was the most significant in random forest classification，with F1-score and AUC reaching O.96 and AUC，respectively.Combined with Mueler matrix polarization decomposition method and texture analysis，the proposed multivariate error correction model can effectively distinguish papilary thyroid carcinoma and nodular goiter samples，which provides a new method for early auxiliary diagnosis of cancer and has a good application prospect.

Key words： polarization imaging；error correction； Mueller matrix vector parameters； texture features;random forest

1引言

研究表明，恶性肿瘤防控问题成为当前“健康中国"建设中的重要公共卫生挑战之一[1]。（剩余20740字）