Deep learning based identification and classification of cancer cells in cerebrospinal fluid samples

• Applicant: Privatdozentin Dr. Leonille Schweizer
• Subject Area: Molecular and Cellular Neurology and Neuropathology
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 545407862

The analysis of cerebrospinal fluid is a cornerstone in the clinical workup of patients with neurological symptoms for the differentiation of inflammatory, neoplastic or hemorrhagic processes. Cytology is an integral and indispensable part of cerebrospinal fluid analysis and tool for disease and treatment monitoring. Detection of neoplastic meningitis is of particular importance for oncological patients and treating physicians to define the prognosis and optimal therapeutic strategy. Gold-standard morphological profiling and manual counting of CSF cells relies on visual inspection by highly-specialized technicians and/or board-certified neuropathologists. Cancer cell detection in CSF samples not only requires differentiating between normal and abnormal cells, but also usually relies on simultaneously assessing the origin of the neoplastic cells, which is especially relevant for patients with cancers of unknown primary. Cytometers are unable to detect pathological cell types, such as tumor cells or mitoses. Therefore, cytometer-based cell differentiation is currently not recommended and labor-intense and time-consuming microscopic evaluation of each sample is still indispensable to avoid misdiagnoses. In our recent project, we trained a convolutional neural network capable to identify 15 diagnostically relevant cell types, among them tumor cells. The aim of the proposed project is to establish a clinical-grade algorithm and workflow for cancer cell detection and classification in cerebrospinal fluid samples by applying a neural network based image analysis method on digitized cerebrospinal fluid specimens. The current preliminary cancer cell dataset will be extended by further neoplastic meningitis cases covering common and rare cancer types known to disseminate to the central nervous system. The neural network needs to be adopted and improved for cancer cell detection and a subclassifier trained to predict cancer cell origin. Furthermore, the open source segmentation algorithms for automated cell detection have to be optimized for cerebrospinal fluid specimens. Finally, a prospective clinical validation of the algorithm in parallel to routine diagnostics is necessary to compare the cancer cell detection rate and cell of origin prediction to neuropathologists. The implementation of the algorithm in the diagnostic workflow will not only provide the basis for an automated cancer cell screening in routine, but also allow the determination of absolute and relative cell counts with the aim to identify intra-disease heterogeneity and study the intrathecal immune cell profile accompanying various disease, among them neoplastic meningitis.

• DFG Programme: Research Grants