Measurement of placental depth during time domain NIRS using deep learning

MAESTROS is a state-of-the-art in-house developed multi-wavelength time-domain NIRS system. Using NIRS to measure in-vivo placenta oxygenation non-invasively at the bedside could potentially provide valuable insights into the health status of the pregnancy. However, the variable depth of the placenta in the abdomen results in reliability issues for monitoring with NIR. Here, a deep learning model is presented to estimate the placental depth using the Distribution of Time of Flight (DTOF) measurements from the MAESTROS system. The model trained with 108 cases predicted the placental depth in 20 test cases with a mean error of 0.42 cm and a strong statistical correlation between predicted values and the measurements from the ultrasound scans. The model was 100% accurate when identifying the 20% of cases where the placenta is deeper than 3 cm, where the depth is great enough to undermine NIRS. The model could be used to alert TD-NIRS operators early in the acquisition about placental depth or could assist with data cleaning in study analysis. Furthermore, a technique for explainable Artificial Intelligence was applied to provide insight into the features of the DTOF data used by the model to predict placental depth, which were consistent with expectations based on the physics and anatomy of this application.