Visualization of Vasculature by Non-Rigid Registration and Segmentation of Subharmonic Ultrasonic Images

Introduction: Detection of abnormal vasculature in pancreatic cancer tumors may allow early diagnosis, more effective treatment and improve the survival rates of patients. Contrast-enhanced ultrasound (CEUS) is a commonly used imaging modality that promotes visualization of the vasculature of the tissue and may help detect small morphological changes. Non-rigid registration of images (frames) within a clip may be beneficial to overcome some of the problems inherent to the physiological motion of the observed anatomical structures [1]. Segmentation of different types of tissue within the clip, differentiating the dynamics of noisy background (tissue) from that of flow (vascularization), may further improve the visualization of the vasculature. Herein we present improved visibility of the vasculature by utilizing a two-step method. In the first step, a non-rigid registration is performed of the images in the clip acquired by grayscale imaging, and consequently applying the estimated deformation field to the images obtained by the employment of subharmonic imaging (SHI). In the second step, the registered clip is segmented by analyzing the dynamics of vascular flow, as visualized by SHI.

Methods: Data: Two groups of data were used: Group 1 - human pancreas with pancreatic tumor; Group 2 – mice hindlimb with induced tumor. The data were acquired using a curvi-linear transducer with a GE Healthcare research Logiq 9 or E9 system (employing a dual mode of transmissions of 4-MHz for the grayscale imaging and pulse inversion transmission at 2.5-MHz for the SHI, while receiving at 1.25-MHz for SHI) [2]. The pulse-echoes of the fundamental harmonic and subharmonic signals were generated by employing interleaving acquisition mode, and are shown simultaneously. The clips were acquired after a bolus injection (Group 1) or continuous injection of contrast agent (UCA) (Group 2).

Step I - Registration: The Morphon (i.e. a local phase) registration method was used, by applying directional quadrature filters to the grayscale images and estimating the deformation field from the resultant filtration response [3, 4]. This deformation field, which represents the new position of each pixel in the deformed image is then applied to the SHI clip. The results of registration of the clips are presented, as well as maximum intensity projection (MIP) processing applied to the clips preceding/succeeding the registration.

Step II – Flow Segmentation: Local spatial-temporal feature extraction and block processing were employed: selected regions of interest (ROI’s) within the SHI clip were decomposed by wavelet decomposition, for both the noise reduction and the extraction of local spatial-temporal behavior [5]. Since similar tissues have similar properties but may have a shift in phase in the temporal domain, a time-independent discrete cosine transformation (DCT) was performed, followed the spatial-temporal feature extraction. Next, ROI’s representing different properties were selected: 1) background tissue – containing a low concentration of UCA, 2) blood vessels - containing high concentrations of UCA, 3) blood vessels containing low concentrations of UCA. The clustering of these areas was applied to segment the SHI clip [6].

Results: The results, best presented as clips, are here presented and validated by depicting the correlation between frames (before and after the registration). Additionally, MIP images are presented, illustrating the total movements of the organ and the vessels.

The results demonstrate the increase of the correlation. In group 1, a significant improvement of the correlation is depicted after performing registration at the end-expiration period. Further improvement is achieved for the registration of a full breathing cycle. In the MIP images of the unprocessed grayscale and SHI data, the edges of the vessel are smudged due to its movement, as opposed to the MIP result after applying the registration data. An example from Group 2 is given, with higher correlations, since there is less movement in the hindleg tumor clips of the mouse.

Results of the segmentation of different flow patterns, as viewed in the respective SHI clips, are depicted for an example from Group 1 (human pancreas), and from Group 2 (mice tumor). In both cases, capillary flow is not marked, while flow in arterioles and venules is marked in red, and rapid flow in blue.

Discussion and Conclusions: Both the clinical and the pre-clinical data were acquired under conditions that severely limit the ability to register and process the images. First, the acquisition frame rates were very low (7-8 FPS and 18-20 FPS, respectively) affecting the quality of motion estimation when movements occur. Moreover, the grayscale images and SHI were acquired in an interleaving mode, leading to temporal mismatch due to the time gap between acquisitions, thus affecting the correction. In addition, UCA was infused without employing the flash-replenishment technique, thus the UCA was circulating for a long time and filling up capillaries, which makes the segmentation of the blood vessels more challenging. These shortcomings will not be overcome in the near future. In addition, the grayscale clips of the pancreas suffer from a low SNR and significant clutter noise, which interfere with the pre-processing, registration and segmentation processes.

The correlation between the frames of the SHI clips was improved for both groups, indicating an improved alignment. The visualization of the tissue vasculature was also improved after the registration. Segmentation of the vasculature by their properties further enhanced the visualization. The human pancreas data, which commonly suffers from large movements and low FPS, benefits from the two-step registration, e.g. registration during the end-expiration followed by the registration of the entire clip. When only small movements are present, as shown in Fig. 3, there is no need for the two stages solution.

The limitation regarding the data should be further investigated in the future. One potential way of overcoming the impact of these shortcomings is preprocessing before registration, where for clips with large movements local registration should first be performed, and only then global registration. Moreover, a post-segmentation process can additionally be performed by utilizing the statistical properties of UCA groups, by removing the regions that were segmented as having large UCA concentration and reprocessing the regions of low UCA concentration for enhancing the visualization of smaller vasculature.