Background: Difficulty diagnoses of leptomeningeal metastasis and lack of typical symptoms are considerable challenges. It is necessary to establish a deep learning model to classify cancer cells in CSF to facility doctor early diagnosis.

Method: a total of 53255 cancer cells were identified from the cerebrospinal fluid laboratory of Xijing Hospital. Cancer recognizes model (lymphocytes, monocytes, neutrophils, erythrocytes, cancer cells), cancer classification model(lung cancer cells, gastric cancer cells, breast cancer cells, pancreatic cancer cells), and eight-ways classification (lymphocytes, monocytes, neutrophils, erythrocytes, cancer cells）performance of models was evaluated. The generalizability of our models was investigated on data from the validation set or external test set. Final model performance was compared with interpretation from 42 doctors of varying levels of experience using the human-machine tests.

Results: the mAP of cancer recognize model(CNN1) was over 95%, the mAP of the cell classification model (CNN2) was nearly 80%. We successfully achieved cell training in the small cell dataset in an 8-way cell classification model(CNN3), which overall accuracy is 93%. In the human-machine test, the cell classification model accuracy was close to experts and higher than other doctors. The cancer cell classification model's accuracy is 10% higher than that of experts, and the time-consuming is only 1/3 of that of an expert. The 8-ways cancer classification model performance was better in the tests. It saves 90% working time of cytologists with the computer-assisted diagnosis (CAD) software introduce.

Conclusion: Deep learning can be widely used in cerebrospinal fluid cytology pictures 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.