Deep learning-enabled quantification of the capillary network in the myocardium by autofluorescence imaging

Myocardial capillaries are responsible for exchanging gases, nutrients and waste products between blood and cardiomyocytes, playing a critical role in maintaining cardiac function and homeostasis. The capillaries in tissue sections are typically assessed by immunohistochemistry and microscopic examination. However, staining artifacts and unsatisfying signal-to-background can impede the interpretation of capillaries, compromising assessment accuracy. Here, we present a label-free approach (FreeQC) to quantify myocardial capillaries in aldehyde-fixed tissue sections using autofluorescence imaging and deep-learning assisted semantic segmentation. The U-Net architecture was utilized to extract semantic features and produce pixel-level labels of cardiomyocytes and capillaries. Capillary-related measurements were calculated from labeled pixels. Four state-of-the-art U-Net variances were evaluated for their segmentation and quantification performance against ground-truth (human annotation and assessment). Trained on a small-scale dataset containing fewer than 100 images, FreeQC demonstrated the efficacy in performing automated quantification of myocardial capillaries in tissue sections without exogenous labeling.