Differentiation of leukocytes in bronchoalveolar lavage fluid samples using higher harmonic generation microscopy and deep learning

In many diseases such as interstitial lung diseases (ILDs), patient diagnosis relies on diagnostic analysis of bronchoalveolar lavage fluid (BALF) and biopsies. In BALF the differentiation of neutrophils, eosinophils, lymphocytes, and macrophages can contribute to diagnose the underlying ILD entity. To analyze the BALF standard cytological techniques are labor-intensive and time-consuming. Studies have shown promising cell identification performance on blood fractions analyzed by third harmonic generation (THG) and multiphoton excited autofluorescence (MPEF) microscopy. Here, we extend this to BALF samples, and we trained a deep learning algorithm for automated analysis on the image level against reference cytology. We imaged blood fractions from three healthy individuals and one asthma patients, and six BALFs from ILD patients. We determined the leukocyte characteristics in terms of cellular and nuclear morphology, and THG and MPEF signal intensity. A deep learning model was trained on both blood fractions and BALF 2D images was used to estimate the leukocyte ratios by using only the standard cytology differential cell ratios at the image-level as reference. The deep learning network has learned to identify individual cells and was able to provide a reasonable estimate of the leukocyte percentage, coming within a 2 to 10% margin in BALF samples in the hold-out testing set. We suggest that the performance of the combined label free imaging and AI analysis can be improved further by collecting 3D data and data of additional fluid samples of various ILD diseases and healthy samples, THG/MPEF microscopy in combination with deep learning is a promising technique for instant differentiation and quantification of leukocytes. Immediate feedback on leukocyte ratios will not only speed-up the diagnostic process but can also reduce costs, the workload and reduce inter-observer variations.