AUTHOR=Yu Wenjin , Liu Yangyang , Zhao Yunsong , Huang Haofan , Liu Jiahao , Yao Xiaofeng , Li Jingwen , Xie Zhen , Jiang Luyue , Wu Heping , Cao Xinhao , Zhou Jiaming , Guo Yuting , Li Gaoyang , Ren Matthew Xinhu , Quan Yi , Mu Tingmin , Izquierdo Guillermo Ayuso , Zhang Guoxun , Zhao Runze , Zhao Di , Yan Jiangyun , Zhang Haijun , Lv Junchao , Yao Qian , Duan Yan , Zhou Huimin , Liu Tingting , He Ying , Bian Ting , Dai Wen , Huai Jiahui , Wang Xiyuan , He Qian , Gao Yi , Ren Wei , Niu Gang , Zhao Gang 

TITLE=Deep Learning-Based Classification of Cancer Cell in Leptomeningeal Metastasis on Cytomorphologic Features of Cerebrospinal Fluid

JOURNAL=Frontiers in Oncology

VOLUME=Volume 12 - 2022

YEAR=2022

URL=https://www.frontiersin.org/journals/oncology/articles/10.3389/fonc.2022.821594

DOI=10.3389/fonc.2022.821594

ISSN=2234-943X

ABSTRACT=Background: Difficulty diagnoses of leptomeningeal metastasis and lack of typical symptoms are considerable challenges. A gold diagnosis standard of LM is positive CSF, and cytologist consumes plenty of time to classify cells under a microscope. 
Objective: to establish a deep learning model to classify cancer cells in CSF can facilitate doctor early diagnosis LM. 
Method: a total of 53255 cancer cells were identified from the cerebrospinal fluid laboratory of Xijing Hospital. Five-ways cell classification model(CNN1)consists of lymphocytes, monocytes, neutrophils, erythrocytes, cancer cells, four-ways cancer cell classification model(CNN2)consists of lung cancer cells, gastric cancer cells, breast cancer cells, pancreatic cancer cells. CNNs were constructed by Resnet-inception-V2, we evaluated performance of models on the validation set or test set. Model performance have investigated on two external dataset and compared with 42 doctors of varying levels of experience in the human-machine tests. We also develop Computer-assisted diagnosis (CAD) software to get cytology diagnosis reports rapidly in  research.
Results: On the validation set, the mAP of CNN1 was over 95% and the mAP of CNN2 was nearly 80%. CNNs is successful to classify cells in CSF to facilitate screen the cancer cell. In the human-machine test, CNN1 accuracy was close to experts and higher than other doctors, while CNN2 accuracy is 10% higher than that of experts, and the time-consuming is only 1/3 of that of an expert. It saves 90% working time of cytologists with CAD software. 
Conclusion: A deep learning method which can effectively utilize labeled data through step-by-step training has been developed. Our research constructs the dataset containing over 50000 cells in CSF and successfully classified cancer cells in the CSF to diagnose LM early. Meanwhile, this unique research predicts cancer's primary source of LM relies on cytomorphologic features without immunohistochemistry. Our results showed that deep learning can be widely used in cerebrospinal fluid cytology images for classification. For complex cancer classification tasks, CNN accuracy is significantly higher than that of specialist doctors, and its performance is better than that of junior doctors and interns. With the introduce CNNs and CAD software, it is taken to lighten the load of cytologist.