Standardization of blood flow measurements by automated vascular analysis from power Doppler ultrasound scan

Purpose

Power Doppler ultrasound imaging provides a non-invasive method to explore tissue vascularity in real-time. This has far-reaching clinical utility, for example to assess the perfusion of organs such as the placenta. The power Doppler is less affected by the angle of insonation than color Doppler and sensitive to multiple directions of flow and low flow velocities, therefore more useful to assess tissue perfusion. However, like all forms of ultrasound, power Doppler signal is attenuated by the tissue that it passes through. To compensate for the attenuation and permit the interpatient comparison, it is necessary to standardize the power Doppler signal. The only validated method for estimating tissue perfusion from a 3D ultrasound volume is fractional moving blood volume (FMBV). This requires identification of a large reference vessel close to the target of interest which can be assumed to have 100% vascularity. This vessel is then used as a standardization point to calculate a quantitative estimate of perfusion, FMBV. This can be performed manually but to be clinically useful as a screening test, the calculation of FMBV needs to be fully automated.

The objective of this study is to propose a fully automated method for the estimation of FMBV and demonstrate its application in the human placenta by using the 3D power Doppler ultrasound scans of the first trimester placenta.

Method description

The 3D power Doppler ultrasound data accompanied with B-mode ultrasound scans was collected as part of a research study conducted with full local ethical approval [1]. The scans were acquired in the first-trimester of pregnancy by a GE Voluson E8 scanner and a 3D/4D curved-array abdominal transducer. A coordinate transformation was performed to convert the raw data in Kretz file format using a toroidal system to Cartesian coordinates for the application of further image processing [2]. After coordinate transformation, an automated processing pipeline was applied to the data which consists of (I) vascular geometry analysis, (II) placental segmentation and representative vessel segment selection, and (II) FMBV estimation by standardization of the power Doppler signal.

(I). Vascular geometry analysis

Initially, a multi-seed region growing based method was applied to segment all the blood vessels from the 3D power Doppler scans (Fig. 1 (a) and (b)). Next, a medial axis based thinning method was used to extract the vascular skeleton. The vessels were then separated into individual segments lying between the bifurcations using the watershed technique with local vessel centerlines as markers (Fig. 1 (c)). The signal intensity gradient information was then used to identify individual vessel segments within the complex vascular structure.

Fig.1.

Scheme of the automated 3D FMBV estimation. (a) power Doppler signals overlaid on B-mode ultrasound scans; (b) blood vessel overlaid on a B-mode ultrasound scan, detected surrounding the placenta (in yellow); (c) vessel centerline and the segments split; (d) reference vessel selected in the standardization VOI (in blue) and the corresponding cumulative curve of the power Doppler signals within this reference vessel.

(II). Placental segmentation and reference vessel selection

The utero-placental interface (UPI) where the maternal blood enters the placenta, is the site of abnormal vascularity in pregnancies destined to develop many adverse outcomes. A fully convolutional neural network was applied to the B-mode ultrasound volumes to segment the placenta, amniotic fluid and fetus using a multi-class model. A total volume combining these volumes was generated. The UPI is the interface between this volume and the placental volume. An appropriate volume of interest (VOI in blue in Fig. 1 (d)) used as the standardization volume was then defined relative to UPI. The largest individual vessel within that standardization VOI was identified as the reference vessel (cyan segment marked by a white circle in Fig. 1(d)).

(III). FMBV estimation

The power Doppler signal intensities of the reference vessel selected within the standardization VOI were used to build the cumulative probability distribution function (PDF). According to the method proposed by Rubin et al [3], the knee value of the cumulative PDF was used as a standardization value (Fig. 1 (d)).

The target VOI within the placenta (0.5 cm from the UPI into the placenta shown in green in Fig. 1(d)) was normalized by comparing the voxel intensities with the standardization value computed from the reference vessel VOI. The resulting FMBV value was the mean of the normalized power Doppler values expressed as a percentage related to 100% vascularity.

Results

The proposed method was applied to 3D power Doppler image data with promising results. These appeared to indicate a good performance on the blood vessel segmentation and skeletonization as the vessels identified were all continuous without any outlying segments. The appearance was consistent with what was subjectively judged to be representative of the vascular plexus by an experienced sonographer. Vessel cross-sections were extracted for the evaluation of the gradient based identification of individual vessel segments. These cross-sections were classified manually into two groups by counting the number of signal intensity peaks regarded as ground truth. The proposed gradient based method was applied to these vessel sections distinguishing the individual vessel well.

The multi-class convolutional neural network obtained a Dice similarity coefficient (DSC) of 0.81 for placenta segmentation in 141 patient data. This permitted the accurate localization of the target placental volume and the reference blood vessel being use as the standardization point for estimation of FMBV.

The automated FMBV estimation method was performed on 20 patient data. The 3D FMBV value was 21.35 ± 9.43 (mean ± STD). This served as an efficient tool for assessing the blood perfusion in placenta bed.

Conclusion (including the ‘New or breakthrough work to be presented’)

In this study, a fully automated method was proposed to standardize the power Doppler signals to estimate tissue perfusion known as fractional moving blood volume (FMBV). Promising results were obtained from the 3D power Doppler scans. Due to the signal sensitivity to noise ratio in small vessels and the sensitivity to the motion of tissue, accurate mapping of complex vasculature using power Doppler is challenging. To the best of our knowledge, this is the first successful attempt to automatically identify individual blood vessel segments from a complex vascular plexus as the ‘standardization’ point for estimation of tissue perfusion with FMBV. The multi-seed region growing based segmentation of the blood vessels is superior to the common used thresholding method [4–6] in maintaining of the vessel connection. The FCNN provided state-of-the-art segmentation of uterus structures for standardization and target VOIs localization. The automated procedure which has the potential to revolutionise the ability of ultrasound to produce vascular measurement not only within the placental bed but also in other organs and tumors.