Development And Assessment Of Artificial Intelligence Models For Intraoperative Pathological Diagnosis Of Pulmonary Lesions

Purpose:Compared to frozen sections,cytological smear gives better preservation of morphological detail and clarity.Yet,intraoperative cytology is complex and requires the collaboration of an expert cytopathologist and surgeon.In recent years,Convolutional neural networks(CNN)have been utilized for tumor diagnosis and tumor histological staging cytology,and it has been demonstrated that the use of software to analyze the cytoarchitecture in digitized cell block material is more reliable than the visual assessment of some cytopathologists.In this study,CNN is used to develop a framework for feature extraction-based classification models to model intraoperative cytopathological diagnosis of pulmonary occupational lesions,to compare the differences between AI interpretation and human interpretation,and to assess the clinical utility of the models.Methods:During 2021.02 to 2022.04,a total of 368 instances of pulmonary occupational lesions and 405 lesions were diagnosed at the China-Japan Friendship Hospital.Using intraoperative samples of fresh lung tissue,cytological smears were created.The specimens were viewed under a microscope after fixation and staining.The cytological smear photos were obtained with a digital camera at 20x magnification,and two experienced pathologists collaborated to identify the cancerous regions.A neural network developed with convolutional layers and trained to produce a feature map that can discriminate malignant photos from non-malignant images scales the original images and extracts features.The one-dimensional feature vector created by unfolding the feature map output of a CNN is then used for classification training in a multilayer perceptron.Results:In the postoperative paraffin pathological diagnosis of 16 categories of pulmonary lesions,89.4%of the 405 lesions were determined to be malignant while 10.6%were determined to be benign.The experiment included a total of 610 cytological smears,of which 57.2%were benign,40%were malignant,and 2.8%were unclear.For the independent test set,the framework of the feature extraction-based classification model created in this study achieved an accuracy of 0.94,a sensitivity of 0.97,a specificity of 0.93,a kappa value of 0.83,and an AUC of 0.98.Comparatively,the precision of primary and advanced cytopathologists was 0.92 and 0.97 respectively.To validate the generalizability of the model,it was applied to images of pleural effusion and bronchoalveolar lavage,yielding kappa values of 0.88 and 0.78 and AUCs of 0.93 and 0.08,respectively,demonstrating that the framework had good generalizability.Conclusions:Intraoperative cytological smear testing provides a speedy,efficient,and cost-effective procedure for intraoperative pathological diagnosis with a high application value,particularly for primary hospitals without the infrastructure to perform frozen section pathological diagnosis.The model developed in this study demonstrated good accuracy,as well as high generalization ability and robustness,and can be used for preliminary intraoperative cytological pathological diagnosis of the lung,as well as in primary hospitals to compensate for a lack of experienced cytopathologists.Purpose:In this study,we apply the previously developed lung cancer histological classification system to frozen section diagnosis,and propose an optimization scheme for the model,expecting to achieve accurate recognition and accurate histological typing of lung cancer on frozen sections.Methods:Frozen sections of benign and malignant pulmonary occupational lesions from the China-Japan Friendship Hospital were digitized and input into the model,and the results of lung cancer identification and pathological typing were output.We optimized the model by expanding the training data set,adjusting the model structure and using double-threshold strategy,and then retested the model to evaluate the optimization and the feasibility of applying it to frozen section diagnosis.Results:The AUC,accuracy,sensitivity and specificity of the pre-optimisation model were 0.536,0.81 and 0.34.After two optimizations,the AUC,accuracy,sensitivity and specificity of the model were 0.67 and 0.80,0.68 and 0.74,0.81 and 0.81,and 0.54 and 0.66.In comparison with the pathologist’s observations based on HE images,the model demonstrates good recognition of malignant lesions on frozen sections and has the ability to detect unfamiliar components and prompt the pathologist through thermographs,preventing the occurrence of missed diagnoses.After optimization,the model showed a significant improvement in the recognition of benign lesions,especially granulomas,and some improvement of dense lymphocytes areas.The AUC of the model was improved to 0.87 by using double-threshold strategy.In addition,the model also performed well in the pathological typing of lung cancer on frozen sections.Conclusions:In this study,we applied the pre-developed lung cancer histological classification system to the diagnosis of frozen sections of pulmonary occupational lesions,and then optimized and adjusted the model,and initially validated the effectiveness in the diagnosis of frozen section pathology.