Abstract 716P
Background
To develop a stain-free, high-throughput technique for identifying lung cancer metastasis cells in cerebrospinal fluid based on diffraction imaging and deep learning techniques.
Methods
Several lung cancer cell lines prone to brain metastasis were selected and subcultured with F1640 culture medium plus 10% fetal bovine serum in a 37° constant-temperature incubator containing 5% CO2. Before use, the cells were digested with 0.25% trypsin and centrifuged at 1000 RPM for 5 minutes. After centrifugation, the cells were collected and re-suspended in phosphate buffer (PBS), and the cell concentration was adjusted to 1×107/ml. Whole blood from healthy volunteers was collected. Red blood cells in these samples were lysed, and then samples were centrifuged at 450g for 10 minutes to collect other cells. The cells were then re-suspended in PBS to 1×107/ml. Diffraction images of the cells were acquired using polarization diffraction imaging flow cytometry with a TDI (Time Delay Integration) camera. A pair of images (S-polarized diffraction patterns and P-polarized diffraction patterns) was obtained for each cell. The image pairs were divided into training and test data sets by random sampling, accounting for 75% and 25% of the total data. The in-house developed automatic identification model of lung cancer metastasis cells using deep learning techniques consists of 3 modules (data preprocessing, cell identification, and data training). A convolutional neural network was established based on Visual Geometry Group, which has 10 convolutional layers divided into 5 blocks.
Results
The automatic recognition model was trained with the training data sets and tested on the independent testing data sets. The results show that the model can achieve a high identification accuracy of 0.96, with sensitivity and specificity of 0.98 and 0.96, respectively.
Conclusions
The model with a convolutional neural network and densely connected classifiers can successfully identify lung cancer cells. Compared with cell detection techniques using fluorescent staining and antibody labeling, the operation process with the new model is simplified, and the time and reagent costs are reduced.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
The authors.
Funding
Natural Science Foundation of China funding No. 82103691.
Disclosure
All authors have declared no conflicts of interest.