Experimental evaluation of a 3-D fully convolutional network for learning blood oxygenation saturation using photoacoustic imaging

Deep learning approaches can be implemented to estimating blood oxygen saturation (sO2) using spectroscopic photoacoustic imaging without reference to a physical model and thus can be more accommodating of incomplete knowledge of the relevant imaging physics. However, they pose a challenge, namely the requirement for high quality training data with an accurate ground truth. The challenge is particularly pressing for convolutional networks that use photoacoustic images as inputs to learn sO2 because they require high quality simulated training images that accurately replicate these features and this represents a non-trivial challenge. The aim of the study was to assess the scale of this challenge using a convolutional network that used 3D multiwavelength photoacoustic images as inputs.