Automated counts of flowing single red blood cells imaged in living mouse retinal capillaries

Abstract

Purpose : While single red blood cells (RBCs) have been imaged in capillaries with adaptive optics scanning light ophthalmoscopy (AOSLO), quantification of flux has required manual determination which is impractical as capillaries can pass >30,000 cells per minute. Here, we report a fully automated strategy based on a convolutional neural network (CNN) to rapidly count RBCs flowing in the same capillary over hours.

Methods : Two anesthetized healthy C57BL/6J mice were imaged with a custom-built AOSLO using near infrared phase contrast. 15kHz line scan imaging across single capillaries produced images of discrete RBCs as they travelled across the imaging beam. To train and validate the algorithm, RBC positions were first manually determined by 3 graders. Ground truth RBCs were determined when ≥ 2 out of 3 graders marked a cell. A 15-layer CNN was built and trained on 660,555 patches (each 32x32 pixels) with 13,770 manually identified RBCs in a single capillary of one mouse. 3 features were classified (background tissue, plasma gap and RBC). A capillary from a second mouse was imaged semi-continuously 1.8 hrs. 1 min of data was analyzed every 4 minutes with the trained algorithm. Performance was compared to human counts from the first 5 s of data.

Results : The algorithm showed a strong match with human graders with 0.981 sensitivity and 0.975 precision (9 false positives, 7 false negatives in 361 ground truth RBCs, Fig.1). It showed a 4x improvement in speed compared to human graders. 113,187 RBCs were detected over 20 intervals analyzed (Fig 2). At the scale of seconds, variability was dominated by RBC aggregations and spontaneous patterns of flow. At minutes, Fourier analysis of flux data shows a strong influence of cardiac pulsatile flow. A heterogeneous pattern of RBC flow was observed across 1.8 hrs within the same capillary. When comparing mouse heart rate and RBC flux measured over 1.8 hrs, we found a weak positive correlation indicating a potential dependency of retinal capillary flux on systemic cardiac output.

Conclusions : Our fully automated algorithm accurately and precisely counts RBCs comparable to the agreement between human graders, mitigating human fatigue while improving analysis speed on large datasets. This strategy will enable study of RBC dynamics across long periods of time in the same capillaries in conditions of health and disease.

This is a 2020 ARVO Annual Meeting abstract.

 

Figure 1

 

Figure 2