Automatic photoacoustic monitoring of perinatal brain hypoxia with superior sagittal sinus detection

Abstract

Significance: Despite advances in perinatal medicine over decades, perinatal hypoxic-ischemic encephalopathy (HIE) remains a significant cause of fetal cerebral palsy and can lead to other severe medical sequelae or death. Therefore, it is highly desirable to effectively detect brain hypoxia during labor and postnatally for HIE management.

Aim: We recently validated the feasibility of transcranial photoacoustic (PA) imaging for oxyhemoglobin saturation measurement at the superior sagittal sinus ( $O_2{\mathrm{Sat}}_{\mathrm{ss}}$ ) in the neonatal piglet brain, at which overall oxygen supply status can be reflected as a primary collective vein. We aim to automate the PA-based workflow of at-risk subject detection and enable fully autonomous and continuous perinatal monitoring.

Approach: We proposed a two-step algorithm that focuses on the most informative region of the brain for oxygenation status, the superior sagittal sinus (SSS). First, a convolutional neural network (U-Net) is trained to detect the location of SSS in the coronal cross-section PA images. Then, an optimized region of interest patch around the predicted SSS location is cropped from the spectral unmixed image and averaged as the $O_2{\mathrm{Sat}}_{\mathrm{ss}}$ measurement. A confidence score can be computed for the measurement via Monte Carlo dropout (MCD), which infers the prediction uncertainty for better clinical decision-making.

Results: The algorithm was evaluated on an in vivo piglet brain imaging dataset containing 84 independent experimental settings from 10 piglet subjects. A 10-fold leave-one-subject-out cross-validation experiment reports 85.2% sensitivity and 93.3% specificity for healthy/hypoxia classification with an $R$ -squared value of 0.708 and a confidence score of 94.06% based on MCD computation, well agreed with our ground-truth given by blood gas measurements.

Conclusions: The proposed automatic $O_2{\mathrm{Sat}}_{\mathrm{ss}}$ monitoring solution demonstrated a hypoxia detection capability comparable to the human expert manual annotation on the same task. We concluded with high feasibility for a noninvasive PA-based continuous monitoring of the perinatal brain.

Keywords: brain monitoring; deep learning; oxygen saturation measurement; perinatal health; photoacoustic imaging.

Fig. 1

In vivo piglet experiment with controlled oxygenation and multiwavelength photoacoustic imaging. (a) Experiment setup with oxygen input controls and multimodal measurements. (b) 21-wavelength photoacoustic imaging example where brain and skin surface structure are clearly visible with appropriate dynamic range settings.

Fig. 2

Dataset overview. For different piglet subjects, large intensity and distribution variations can be seen: the top row shows images with a dynamic range min-max normalized. For different ${\mathrm{FiO}}_2$ data points of the same piglet subject, the raw photoacoustic spectrum images are visually similar unless spectral unmixing is performed.

Fig. 3

Overview of the proposed algorithm workflow: the anatomical landmark of SSS is first identified, and region of interest (ROI) patch around the SSS in the spectral unmixed ${\mathrm{O}}_2\mathrm{Sat}$ map is used as oxygen level prediction.

Fig. 4

Raw PA image dynamic range rectification to enhance the structural information present in the raw PA signals, such that maximum intensity is saturated at spectral average intensity plus $k$ standard deviation of full spectrum intensities, where $k$ is empirically chosen as 0.5 during test time, and varying between 0 and 1 for training augmentation. Example images from each of the 10 piglet subjects are presented.

Fig. 5

ROI optimization for SSS representation. (a) Annotated/predicted SSS location and the optimized ROI representation overlaid on the original PA image. (b) ROI parameterization for both size and displacement. (c) ${\mathrm{O}}_2\mathrm{Sat}$-map-based oxygen level reading compared with blood-sample-based reading using optimized ROI.

Fig. 6

End-to-end learning workflow for ${\mathrm{O}}_2\mathrm{Sat}$ estimation. M1 to M4 corresponds to methods with different inputs and different CNNs. The CNN model for M1 to M3 remains the same except for the input, and a smaller CNN model for M4 is used due to the smaller input size.

Fig. 7

SSS localization error distribution. (a) Zoom-in view of the SSS location prediction compared with the expert annotation for all 84 data points. (b) The predictions and expert annotation overlaid on one example PA image for scale comparison (expert annotation location is from the same example image).

Fig. 8

${\mathrm{O}}_2{\mathrm{Sat}}_{\mathrm{ss}}$ prediction performance quantification with cross-validation experiments over the entire dataset. (a)–(f) Experimental results with different workflows: The proposed method stands for the two-step SSS-localization/spectral unmixing approach, M1 to M4 represents the end-to-end learning approach with different inputs and GT is the expert SSS annotation with ROI spectral unmixing. (g) Classification and regression metrics for different ${\mathrm{O}}_2{\mathrm{Sat}}_{\mathrm{ss}}$ prediction workflows.

Fig. 9

Uncertainty quantification with Monte Carlo dropout (MCD). (a), (d) SSS localization distribution example with model-MCD overlaid on the input raw PA image and its zoom-in view. The opaqueness of the red dots indicates multiple occurrences at the same locations. (b), (e) The distribution example with data MCD. (c), (f) The distribution example with combined (model + data) MCD. (g) Mean STD (mSTD) for different types of uncertainties over the entire dataset, and sO2 stands for the ${\mathrm{O}}_2{\mathrm{Sat}}_{\mathrm{ss}}$ prediction.

Fig. 10

Excluded piglet subjects shown with rectified dynamic range for image plotting. When the PA transducer is placed correctly, i.e., perpendicular to the skull surface above the SSS region, the anatomical structure will follow a general pattern as shown in the bottom right figure with “M” shape brain curvature. (a)–(e) corresponds to the acquired SSS coronal plane PA images of the five excluded piglet subjects A to E, respectively. (f) is the simplified sketch for the brain anatomical structure in coronal view.

Fig. 11

Excluded piglet subjects with abnormal ground truth ${\mathrm{O}}_2{\mathrm{Sat}}_{\mathrm{ss}}$ readings when compared with other piglet subjects in the dataset.