DeepPerfusion: A comprehensible two-branched deep learning architecture for high-precision blood volume pulse extraction based on imaging photoplethysmography

Abstract

Imaging photoplethysmography (iPPG) is a contactless approach for the extraction of the blood volume pulsation (BVP). Analyzing the small intensity changes resulting from fluctuations in light absorption in upper skin layers enables BVP extraction. Inhomogeneous illumination or head movements impede iPPG-based BVP extraction. To mitigate these influences, an important step is the accurate skin segmentation and weighting, which has received insufficient attention in state-of-the-art (SOTA) deep learning-based approaches in particular. Therefore, we propose DeepPerfusion, a two-branched deep learning architecture, that combines precise skin segmentation and weighting as well as BVP extraction into one model. Together with our newly developed patch-based temporal normalization mechanism and our innovative training pipeline, DeepPerfusion achieved highly accurate BVP extraction. We performed a thorough performance analysis and evaluated the mean absolute error (MAE) for heart rate extraction and the signal-to-noise ratio (SNR) on 156 subjects from three publicly available datasets and compared it with nine SOTA approaches that underwent the same training and evaluation pipeline. For the median across subjects of each dataset, DeepPerfusion consistently achieved MAE below 1 beat per minute, outperforming the best SOTA approaches by up to 49 %. Furthermore, DeepPerfusion achieved high SNR with at least 5.81 dB which was about two to three times higher compared to the best SOTA approaches. In contrast to SOTA approaches, DeepPerfusion’s performance was consistent, robust and highly precise. This demonstrates DeepPerfusion’s ability to perform high-precision BVP extraction. We expect this to open up new diagnostic applications for iPPG in the future.