Analysing cerebrospinal fluid with explainable deep learning: From diagnostics to insights.

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Cerebrospinal fluid (CSF) analysis is crucial for diagnosing neurological diseases.
• Current cell typing methods rely on subjective, time-consuming microscopic examination.

Purpose of the Study:

• To develop an automated image analysis approach for CSF cell classification.
• To train a multiclass convolutional neural network (CNN) for differential cell typing in CSF.

Main Methods:

• Expert annotations of 123,181 digitized CSF objects from 78 patients.
• Training a multiclass CNN on 15 clinically relevant cell categories.
• Validation using explainable AI (XAI) and comparison with board-certified neuropathologists.

Main Results:

• CNN achieved high accuracy (mean AUC 97.3%) in classifying 15 CSF cell categories.
• CNN's inter-rater agreement (Krippendorff's alpha 0.79) was non-inferior to human experts (mean 0.72).
• CNN correctly diagnosed 10/11 clinical samples, comparable to human raters (7-11/11).

Conclusions:

• The developed AI approach overcomes limitations of manual CSF cell classification.
• Explainable AI connects machine decisions to cellular properties, offering a versatile diagnostic tool.
• This quantitative method aids in investigating CSF manifestations of neurological diseases.