Bronchoalveolar Lavage as a Candidate for Artificial Intelligence Integration: Insights into Differential Cell Count...

Bronchoalveolar lavage (BAL) brings an important contribution in diagnosing pulmonary diseases. The analysis of standard cell distribution and the assessment of iron-laden macrophages (the Golde score) are integral to standard medical reports. However, the traditional cytological method of manual cell counting is subject to interobserver variability and staining quality issues.

To address these issues, we trained AI-based algorithms to enhance the accuracy of differentiated cell counts of macrophages, lymphocytes, neutrophils, eosinophils, ciliated cells, and squamous cells as well as the Golde score, which assesses the hemosiderin content in macrophages. For this purpose, we assembled an internal sample cohort with 16 Hemacolor, 16 Papanicolaou, and 5 iron-stained smears. For validation, we used 10 slides each of Papanicolaou and Hemacolor staining and 5 with iron staining.

The algorithm achieved fair to excellent correlation compared to two cytologists: For Papanicolaou staining, the correlations were macrophages 0.96, lymphocytes 0.98, neutrophil granulocytes 0.99, eosinophils 0.58, ciliated cells 0.61, squamous cells 0.31. In Hemacolor staining the correlations were macrophages 0.97, lymphocytes 0.92, neutrophils 0.99, eosinophils 0.99, ciliated cells 0.58, squamous cells -0.145. The automated Golde score calculation deviated on average by 19 points from the manual evaluation.

The study demonstrates the potential of AI-supported methods for BAL analysis in diagnostic cytology. The high accuracy in recognizing cell types and calculating the Golde score underlines the benefits of expanding the training data for broader clinical applications. Further research is encouraged to support the use of digital cytology on conventional smears in clinical practice.