Harmonic Motion Imaging guided Focused Ultrasound (HMIgFUS) ablation: comparison between three FUS interference filtering methods

Abstract

Objective

Harmonic Motion Imaging (HMI) is an acoustic radiation-force (RF) based elasticity imaging technique which can be used to monitor changes in tissue mechanical properties due to FUS-induced thermal ablation. In conventional HMI, the amplitude modulated (AM) FUS sequence and imaging pulse are transmitted simultaneously. With this method, the high-amplitude FUS signal must be separated from the imaging data for tissue displacement estimation. Frequency domain notch and bandpass filtering were previously used to filter FUS interference from imaging data. However, FUS interference becomes more pronounced at increased intensities and durations necessary for thermal ablation, rendering frequency domain filtering challenging.

In this study, three methods were compared for FUS interference filtering during HMI guided FUS ablation (HMIgFUS). The methods consisted of notch filtering; interleaved HMIgFUS, with interleaved FUS and imaging pulses; and FUS-net, a CNN-based UNET autoencoder developed by our group for FUS interference filtering in radio frequency (RF) data. FUS-net is applied here for the first time for the purpose of ablation displacement monitoring.

Methods

The three filtering methods were tested during 20 HMIgFUS acquisitions in an ex vivo canine liver using a range of peak positive pressures from 11-18 MPa and durations ranging from 60 to 180 seconds. Bmode mean squared error (MSE) and displacement amplitude contrast-to-noise ratio (CNR) was calculated and compared between the three methods.

Discussion

The interleaved method for HMIgFUS was shown to be significantly robust in avoiding FUS interference in all tested cases for FUS ablation monitoring, especially cases with high FUS pressure and long durations, as opposed to traditional notch filtering and FUS-net filtering. CNR obtained from displacement amplitude maps in the interleaved dataset was significantly higher in all cases than that obtained from the notch filtered and FUS-net datasets. There was not a signficant trend in displacement CNR between the FUS-net and notch filtered datasets. However, bmode MSE was found to be significantly higher when comparing the FUS-net and interleaved datasets as opposed to the notch filtered and interleaved datasets, suggesting further potential of FUS-net as a FUS interference filtering method.

Conclusion

These findings indicate the robustness of interleaved HMIgFUS in avoiding FUS interference during HMIgFUS monitoring, and the advantages, limitations, and future potential of FUS-net and notch filtering.

Introduction

Focused Ultrasound Surgery (FUS) is a noninvasive treatment procedure in which large amounts of focused ultrasound energy is administered to target tissue, causing localized temperature rise and a lesion of cell necrosis. FUS has been used clinically and/or tested in numerous studies for treating local diseases in organs such as the breast, thyroid, prostate, and pancreas.^1,2,3,4 Further clinical adoption of FUS relies heavily on the quality of treatment monitoring. To ensure effective ablation and minimize side effects, real-time treatment monitoring is necessary to track the tissue response to FUS, enabling clinicians to prescribe the appropriate FUS dosage and locate lesion formation.

Several imaging techniques are currently being investigated for FUS monitoring. Magnetic Resonance Imaging (MRI) thermometry is used to detect temperature rises from FUS thermal ablation.^5,6,7 However, MRI guidance can be expensive and time-consuming compared to ultrasound-based FUS guidance.^8,9 Comparatively, ultrasound imaging is low-cost and has high temporal resolution.

Of the various ultrasound guidance techniques, ultrasound elasticity imaging is an attractive candidate as its tracking is based on tissue stiffness changes. Since motion tracking follows changes in mechanical properties of the tissue, ultrasound elasticity imaging is not reliant on cavitation or bubble-formation, such as with conventional B-mode hyperecho tracking or passive acoustic mapping.¹⁰ Cavitation or bubble-formation is not directly indicative of the degree of ablation and can impede lesion localization, while tissue stiffness changes have been shown to correlate closely with temperature change and the presence and area of cell necrosis.^11,12

All ultrasound-based motion tracking, however, can be impeded by strong interference from the high intensity FUS signal. During simultaneous application of FUS and imaging, the ultrasound imaging transducer receives both FUS and imaging signals. Since FUS interference generally has a higher amplitude than the imaging signal, image quality is severely degraded.¹³

Harmonic Motion Imaging (HMI) is an ultrasound elasticity imaging technique that enables simultaneous generation and tracking of amplitude-modulated (AM) radiation force induced tissue motion using a focused ultrasound and imaging transducer, respectively.^{14,15,16,17,18} An advantage of HMI is in its transmission of sinusoidal AM radiation force waves, which cause a subsequent sinusoidal tissue displacement at a frequency twice the AM frequency.¹⁹ Since displacement occurs at a specified frequency, HMI displacement can be filtered through frequency-domain filters, removing extraneous noise. Our group has also shown that HMI can monitor FUS treatment response, in that HMI focal displacement changes as a result of heating and subsequent stiffening of the thermally ablated lesion.²⁰ We have shown that our FUS treatment guidance method, termed HMI guided FUS (HMIgFUS), has been shown to effectively locate and monitor the onset and progression of lesion formation in in vitro canine liver^11,21,22,23, in vivo mouse pancreatic tumor^11,24, and ex vivo human breast tissue.²⁵

In conventional HMI, imaging is performed concurrently with FUS application without interruption, such that FUS interference must be removed from RF data prior to displacement estimation. Previous studies in our group have shown interference removal through frequency domain notch filtering, in which the center frequency and harmonics of the FUS transducer are filtered out of received RF data.^11,21,25,26 However, this filtering method is suitable only when the frequency spectra of the imaging and FUS transducers are non-overlapping. Various focused ultrasound and imaging transducer configurations with overlapping frequency spectra may thus make frequency domain filtering unsuitable. Additionally, during FUS ablation, changes in temperature and speed of sound lead to non-linear effects, which increases the energy of higher order harmonics.^27,28 In a setup using a high frequency imaging transducer, which would allow for higher resolution imaging, advantageous especially for small animal studies, and durations and intensities of FUS suitable for ablation, these effects would significantly affect the received frequency spectrum of the imaging transducer, and complicate frequency domain filtering.

As an alternative to frequency-domain filtering, our group has recently developed FUS-net, a method that incorporates a CNN-based U-net autoencoder trained on clean and corrupted RF data in Tensorflow 2.3 for FUS artifact removal.¹³ FUS-net learns the representation of RF data and FUS artifacts in latent space so that the output of corrupted RF input is clean RF data. We have previously shown that B-mode images beamformed from FUS-net RF data shows superior speckle quality and better contrast-to-noise ratio (CNR) in comparison to notch-filtered and adaptive least means squares filtered RF data. FUS-net also outperformed stacked autoencoders (SAE) on test datasets. Preliminary data also showed that displacement estimated from FUS-net filtered data was shown to be similar to notch-filtered data using cross-correlation speckle-tracking algorithms in representative datasets.

Previous datasets used to train and test FUS-net consisted of RF data acquired with a range of peak positive pressures from 0 to 31 MPa peak positive and total FUS treatment durations from 1 to 2 ms, from mouse and human tissues. Frames taken during the FUS pulse were used as corrupted input, while frames taken just before and after the FUS pulse were used as ground truth frames.

The objective of this study was to evaluate three methods of FUS interference removal in HMIgFUS thermal ablation datasets, in which high pressures and long durations of FUS are applied. In such a dataset, additional complications arise in that the tissue is undergoing heating and cavitation. The three methods evaluated were FUS-net filtering; notch filtering, as used in previous HMIgFUS studies; and interleaved HMIgFUS.

Interleaved HMIgFUS is also shown here for the first time. Previous studies have used interleaved sequences for HMI imaging in order to use the same transducer or Verasonics Vantage system to control push and imaging cycles.^29,30,31 This study is the first to use interleaved HMI for the purpose of minimizing FUS interference in HMIgFUS monitoring, while achieving FUS ablation.

The three methods were tested during 20 HMIgFUS acquisitions in an ex vivo canine liver using a range of peak positive pressures from 11-18 MPa and total treatment durations ranging from 60 to 180 seconds.

Materials and Methods

A. HMIgFUS setup

The HMIgFUS setup (Fig. 1.) consisted of a L22-14vxLF linear imaging array with a center frequency of 15.6-MHz (Vermon, Tours, France) confocally aligned with a single-element 4-MHz FUS transducer (Sonic Concepts, Bothell, WA). The imaging transducer was driven by a Verasonics research system (Vantage 256, Verasonics Inc., Kirkland, WA, USA). The FUS transducer was driven by a dual-channel arbitrary waveform function generator (AT33522A; Agilent Technologies Inc., Santa Clara, CA, USA) and amplified by a 50-dB gain power amplifier (325LA, Electronics & Innovations (E&I), Rochester, NY, USA).

Figure 1:

HMIgFUS hardware setup. The pipeline from the workstation to the two transducers is shown on the right, while the transducer setup is shown on the left. The single-element FUS transducer emits a pulse at its geometric focus at the center of the imaging transducer window.

The interleaved FUS signal was generated by the function generator, and consisted of square waves with a 50% duty cycle, with every cycle consisting of 0.5 ms of FUS application followed by 0.5 ms without FUS application. The interleaved FUS signal was designed to approximate a 25 Hz signal, such that the amplitude of each square wave was assigned by its respective position along a 25 Hz sinusoidal wave (Fig. 2).

Figure 2:

Overview of burst FUS and imaging pulse sequences. A. Overview of the square wave burst FUS sequence, showing multiple burst cycles, approximating a 25 Hz sine wave. B. The burst FUS sequence over one half cycle, showing a 50% duty cycle. C. Imaging pulse sequence over-laid on top of the FUS sequence, showing 10 frames of image acquisition per HIFU on-off cycle.

The imaging transducer had a continuous pulse repetition frequency of 10kHz. Tracking was performed during both FUS ‘on’ and ‘off’ cycles. 0.1 seconds of RF data, consisting of 1,000 RF frames, were stored in Verasonics local memory before transferring the data to the workstation. Immediately after this data transfer, the next set of RF frames would be acquired. On average, data transfer took 0.9 seconds, such that RF data was acquired on average once per second during the ablation duration.

The onset of the FUS sequence was synchronized to the onset of the imaging pulses by an analog trigger output of the Verasonics Vantage system to the waveform function generator. From this onset and during the entire treatment duration (60-180 seconds), the interleaved FUS sequence ran continuously. The imaging sequence did not acquire continuously over the entire treatment duration; in order to fulfill memory limitations in the Verasonics Vantage buffer, 1,000 RF frames were acquired and saved at a time. Because this data transfer was not immediate, in order to synchronize the continuous FUS sequence with the non-continuous image acquisition, intensity of the FUS center and harmonic frequencies in RF data was used post-processing to determine the timing of FUS on and off cycles in the image data.

B. Specimen

Ablation was performed in an ex vivo canine liver specimen. Canine liver was chosen due to its optical, acoustic and mechanical properties, and as it is relatively homogeneous, allowing optical delineation of ablated lesions in gross pathology.

A mechanical 3D positioner was used to move the transducers in a 5x4 point raster scan, so that HMIgFUS ablation was performed at 20 locations in the liver specimen. The raster scan step size was 1 cm both in the axial and lateral directions, except in two cases in which the step sizes were 5 mm axially from the previous location, in order to locate a flat surface on the specimen for better coupling. Ablation was performed with a combination of five durations (60, 90, 120, 150, and 180 seconds) and four peak positive pressures (11.25, 13.5, 15.75, and 18 MPa).

C. FUS-net Architecture

FUS-net is a UNET based framework developed previously by our group, trained on clean and noisy RF data for FUS artifact removal. The architecture of FUS-net consists of five symmetric filters (32, 64, 128, 256, 512) used to downsample and upsample the RF data [14]. Each downsampling layer was composed of 2 convolutional layers followed by a batched norm layer. Each upsampling layer receives a concatenated input of the previous layer with its corresponding skip connection to two convolutional layers of the reversed filter order.¹³

D. Data pre-processing

The subsets of image input data to all three filtering methods were taken from the same overall dataset, which consisted of 20 interleaved HMIgFUS treatments. This was performed so that results from each filter method could be directly compared, with the same conditions and tissue properties applied across all datasets. These input data subsets, however, differ in image frame selection at different points of processing (Fig. 3.)

Figure 3:

Flowchart of RF data allocation and processing steps for each of the three datasets: 1. FUS-net (pink), 2. Interleaved (green), and 3. Notch filtering (blue).

For every FUS on-off cycle, 5 RF frames were collected during FUS application, while 5 RF frames were collected following FUS application, for a total of 10 RF frames per cycle (Fig. 3.)

The interleaved dataset consisted of every ninth frame of each FUS square wave cycle, when there was no FUS application. No further preprocessing was done on this dataset.

The FUS-net model was trained using pairs of noisy RF frames (acquired during FUS on) and clean RF frames (acquired during FUS off). Each pair was made up of the fifth frame, during FUS on, and the ninth frame, during FUS off, of one FUS on-off cycle. Thus, for each training pair, the target clean frame was acquired 0.4 ms after the noisy input frame.

Training data was taken from every 10th FUS square wave cycle instead of with every square wave cycle in order to decrease the amount of training data per ablation in order to fulfill memory limitations with the local off-line computing platform, and to decrease training time. With 20 FUS applications with durations ranging from 60 to 180 seconds and on average 1,000 RF frames acquired every second, the total training set consisted of 20,481 pairs of RF data.

A test dataset was compiled consisting of every fourth frame following the start of every FUS on period, such that the test dataset consisted of 213,893 RF frames (Fig. 3.). The test dataset frames did not have any overlap with the FUS-net training set, as in, there were no frames found in both the training set and test set. However, the test dataset came from the same FUS ablation acquisitions as the training set. Both the notch filter method and the trained FUS-net model were tested using this dataset.

RF frames were pre-processed by cropping frames to fit the FUS-net input size of 128x896 pixels, and then scaling the minimum and maximum pixel value between −1 and 1. A train-test split of 80:20, a batch size of 10, and a buffer size of 1,000 were used. FUS-net was trained over 290 epochs. Training and validation for every epoch took on average 30 minutes.

E. Data post-processing and displacement estimation

Interleaved RF data, test dataset RF data, and FUS-net filtered RF data were beamformed using a GPU-based delay-and-sum beamforming method. A 4th order butterworth bandpass filter from 6 to 31 MHz was applied post-beamforming on all datasets.

For the notch filtered dataset, notch filters were implemented on the test dataset RF data following beamforming, in which 2nd order butterworth notch filters with a bandwidth of 1 MHz was applied from the 1st to the 6th harmonic of the FUS center frequency (fc = 4 MHz). Mean squared error (MSE) was calculated pixel-to-pixel between notch filtered and interleaved bmodes, and FUS-net filtered and interleaved bmodes, in order to assess the quality of FUS-net output and compare RF data similarity.

Displacement was estimated from beamformed datasets using a 1D normalized cross-correlation function.³² For each case (interleaved, FUS-net filtered, notch filtered), displacement was calculated from frames separated by 1 ms, for a displacement frame rate of 1kHz. A bandpass filter from 30 to 70 Hz was applied to all displacement data to isolate HMI induced tissue oscillation at 50 Hz.

Peak-to-peak displacement amplitude was calculated per pixel for every 5 cyles of AM displacement. A square 2 by 2 mm region around the focus was used to estimate the mean focal displacement from each displacement amplitude map. The mean focal displacement value at each timepoint was then used to track displacement changes over the ablation duration. Contrast-to-noise ratio (CNR) was calculated as

$$CNR=\mid {\mu }_{foc}-{\mu }_{bkd}\mid ∕\sqrt{({\sigma }_{foc}^2+{\sigma }_{bkd}^2)}$$

with $\mu$ as the mean displacement in a 2 by 2 mm square region, $\sigma$ as standard deviation, $foc$ as a region at the focus, and $bkd$ as a region 4 mm laterally away from the focus (at the same axial depth as $foc$).

CNR in all displacement amplitude maps for each combination of pressure and duration was compared between the datasets. Kruskal-Wallis tests (MATLAB kruskalwallis function) were performed for each of the 20 ablation spots, comparing displacement amplitude map CNR between interleaved, notch filtered, and FUS-net filtered datasets. Tukey-Kramer post-hoc tests were performed on datasets found to be significant following Kruskal-Wallis.

Results and Discussion

Gross pathology revealed clear lesion delineation in at least 14 out of 20 ablation cases, as visually determined from the surface of the liver specimen (Fig. 4). These lesions indicate the presence of thermal ablation. FUS-net training and validation accuracy are shown in Fig. 4, which shows a final train accuracy of 0.999 and a final validation accuracy of 0.93.

Figure 4:

A. Gross pathology of canine liver specimen post ablation, showing the presence of lesioning in locations with 120s of ablation or longer. B. Training (blue) and validation (red) accuracy over 290 epochs.

MSE was found to be significantly lower between the interleaved and FUS-net filtered bmodes than the interleaved and notch-filtered bmodes in all cases, suggesting that FUS-net filtering produced RF data that more closely replicated interleaved data (Fig. 5). This result suggests that FUS-net can reliably produce bmode images that are representative of physiological data, and that do not produce false data.

Figure 5:

A. Boxplots of bmode mean squared error (MSE), comparing interleaved datasets with notch filtered (green) and FUS-net (purple) datasets over each of the 20 ablation conditions. There was a significant difference in MSE in all conditions, shown with an asterisk (kruskal-wallis, p<0.05). B. Boxplots of displacement amplitude map CNR calculated from interleaved (blue), notch filtered (green), and FUS-net (purple) datasets, aggregated for each of the 20 ablation conditions. There was a significant difference in CNR for all ablations (kruskal-wallis, p<0.05). Pairwise significance is indicated with a line and asterisk between groups (tukey-kramer, p<0.05).

Representative images (Fig. 6) show that the interleaved data did not seem to be impacted by FUS interference. Additionally, bmode images filtered with FUS-net lack strong vertical artifacts present in notch filtered datasets. Frequency analysis shows strong similarity in frequency profiles between interleaved and FUS-net filtered beamformed data, in which FUS harmonic frequency amplitudes are reduced, while avoiding complete loss of frequency data as seen in notch filtered data. FUS-net filtered maintenance of frequency information lost in notch filtering contributes to differences in speckle quality. Both of these factors may also contribute to the greater similarity between FUS-net and interleaved bmodes.

Figure 6:

A. Frequency spectrum of beamformed unfiltered, notch filtered, interleaved, and FUS-net data. Data was taken from the center beam in the representative frames shown in Fig.6B, which was 28 seconds after ablation onset with 11.25 MPa peak positive pressure. B. Unfiltered, interleaved, notch filtered, and FUS-net filtered bmodes acquired 28 seconds after ablation onset with 11.25 MPa peak positive pressure. C. Frequency spectrum of beamformed unfiltered, notch filtered, interleaved, and FUS-net data. Data was taken from the center beam in the representative frames shown in Fig.6D, which was 28 seconds after ablation onset with 13.5 MPa peak positive pressure. D. Unfiltered, interleaved, notch filtered, and FUS-net filtered bmodes acquired 28 seconds after ablation onset with 13.5 MPa peak positive pressure.

HMIgFUS displacement estimated using all three processing methods (ground truth, FUS-net filtered, notch-filtered) showed overall similarity in trends in displacement change over treatment duration (Fig. 7). Displacement trends varied across the 20 lesions, with some having rapid increase in displacement and higher overall maximum displacement, while others showed a gradual increase in displacement and lower overall maximum displacement (Fig. 7). This difference may be explained by the presence of acoustically induced cavitation in the rapidly increasing displacement cases, and the lack of cavitation in the gradually increasing displacement lesions. Acoustically induced cavitation is dependent on both rarefaction pressure amplitude and target medium properties.³³ The 20 treated lesions experienced different pressure amplitudes and were applied in different regions of a non-homogeneous ex vivo liver specimen, which may explain why certain treated lesions experienced rapid increases in displacement and potential cavitation, while others did not. The probability of cavitation could also have increased over the 20 lesions due to heating of neighboring tissue, since the 20 lesions were treated sequentially at different locations in a single in vitro liver specimen. Cavitation would cause an increase in displacement by increasing temperature and absorption coefficient at the focal area, thereby increasing radiation force deposition and HMI displacement.^33,34,35

Figure 7:

Mean focal displacement amplitude tracking under specified durations and peak positive pressures, shown in interleaved, notch-filtered, and FUS-net filtered datasets. Mean peak positive focal displacement values were calculated from a square 2x2 mm region around the focus. Shaded regions show standard deviation.

CNR of displacement amplitude maps was calculated in each of the 20 ablation conditions (Fig. 5). CNR was used as a metric to evaluate the relative intensity of the FUS focus in comparison to tissue outside of the focus, as well as noise in both regions. In ideal cases, we assume high displacement at the focal region relative to background tissue, and such assumed that low CNR indicated high standard deviation in these regions from the presence of remaining FUS interference noise. For every ablation condition, there was a significant difference in CNR between the three datasets (kruskal-wallis, p<0.05). In all datasets, the interleaved datasets had significantly higher CNR compared to FUS-net and the notch-filtered datasets (tukey-kramer post-hoc, p<0.05). The FUS-net filtered datasets had signficantly higher CNR than the notch-filtered datasets in seven of the twenty datasets, while notch-filtering also resulted in significantly higher CNR in seven separate cases. In six of the datasets, there was no significance in CNR between the notch and FUS-net filtered datasets.

CNR was also found to be highest at the lowest pressure (11.25 MPa peak positive) across all three filtering methods, and then generally decreased with increasing pressure (Fig. 8) (kruskal-wallis, p<0.05). This suggests that applications of FUS-net and notch filtering may be better suited for FUS applications at lower pressures, with less noise present.

Figure 8:

Boxplots of displacement CNR over pressure, comparing interleaved (blue), notch filtered (green), and FUS-net (purple) datasets. There was a significant difference between all voltages in each of the three datasets (kruskal-wallis, p<0.05).

In cases with gradual displacement change, displacement increase and then decrease could be found in the region near the focus, and FUS-net and notch filtering produced displacement maps that appeared similar to that of the interleaved dataset (Fig. 9). These results, along with Fig. 7, suggest that FUS-net and notch filtering are suitable for use in cases with steady changes in displacement from thermal ablation and low levels of cavitation. In such scenarios, these filtering methods can be used with continuous FUS therapy, which would increase the efficiency of treatment. FUS-net has the additional advantage in these cases, in contrast to notch filtering, of preserving frequency data and bmode quality during FUS application.

Figure 9:

A. HMI displacement amplitude trend over time, for 180 s of ablation with 13.5 MPa peak positive pressure. Displacement at each timepoint was calculated by averaging the displacement amplitude in a 2x2 mm ROI around the geometric focus. Shaded regions show standard deviation. HMI Displacement amplitude maps shown at 0, 60, 120, and 180 s from the same dataset as Fig. 8.A for (B.) interleaved, (C.) notch filtered and (D.) FUS-net datasets. Black squares indicate ROI regions used for Fig. 6.A displacement tracking and CNR calculations.

However, in cases with sharp changes in displacement, FUS-net and notch filtering was not able to adequately filter FUS interference, producing much noisier displacement maps compared to the interleaved dataset (Fig. 10). As shown in Fig. 9, displacement estimated before the onset of the sharp displacement increase could reliably be estimated without significant FUS interference. However, after the sharp displacement increase, FUS-net and notch filtered displacement maps were considerably more noisy. Noise is especially prevalent in regions outside of the focal region, which may explain why tracked displacement trends at the focus were still similar in all three cases despite incomplete removal of FUS noise and lower CNR in FUS-net and notch filtered datasets. In all cases, even in cases with high FUS pressure, the interleaved dataset was the least impacted by FUS interference.

Figure 10:

A. HMI Displacement amplitude trend over time, for 120 s of ablation with 15.75 MPa peak positive pressure. Displacement at each timepoint was calculated by averaging the displacement amplitude in a 2x2 mm ROI around the geometric focus. Shaded regions show standard deviation. HMI Displacement amplitude maps shown at 0, 40, 80, and 120 s from the same dataset as Fig. 9.A for (B.) interleaved, (C.) notch filtered, and (D.) FUS-net datasets. Black squares indicate ROI regions used for Fig. 9.A displacement tracking and CNR calculations.

Notch filtering may not have performed well in cases with high FUS interference because there was too much overlap between imaging and therapeutic frequency spectrums. It was hypothesized that by using clean ground truths without FUS interference, it could be possible to recover clean data from noisy data using FUS-net, in such cases where notch filtering fails. FUS-net filtering obtained higher bmode MSE with interleaved data compared to that of notch filtered data. However, our findings showed similar outcomes between FUS-net and notch filtering datasets in displacement CNR. This suggests that FUS interference was not effectively filtered out by FUS-net filtering in all cases. This issue could arise if the ground truth data in the training datasets were not suitably free from FUS interference, such that FUS-net was not properly trained on isolating FUS interference from tissue speckle. The model could have also been improved by increasing the number of training sets, particularly the number of training sets with sharp increases in displacement and high FUS interference. It may also be possible that FUS-net filtering hindered speckle tracking between frames. The architecture of FUS-net trains on a single RF frame at a time, with no information on adjacent frames, which may hinder FUS interference isolation and speckle preservation. Finally, incomplete FUS-net filtering may be due to a limitation as to how much speckle information the imaging transducer can acquire in the presence of large FUS artifacts, such that clean RF data could not be obtained from noisy RF data with large FUS artifacts.

In this study, within each FUS cycle, the interleaved dataset consisted of RF frames acquired 0.4 ms after the onset of the off portion of the FUS square wave cycle. The dataset was acquired in this manner instead of immediately after the onset of FUS-off because there was found to be some residual interference immediately after FUS was turned off in each cycle. The observed presence of FUS interference within 0-0.3 ms after FUS-off could either impose limits on the duty cycle of the FUS interleaved sequence, such that a longer period of FUS-off is required, or necessitate further filtering of RF data, if a shorter FUS-off duration is desired. In this study, FUS interference within 0-0.3 ms of the FUS-off cycle was found to be much more minimal as compared to the interference during the FUS-on cycle. Thus, this type of minimal FUS interference may be much more simply filtered out using notch filtering or other frequency based methods.

The acquisition of image frames 0.4 ms after the onset of the ‘off’ portion of the FUS cycle for the interleaved dataset could also explain the lower displacement values in the interleaved dataset displacement trends vs that of the FUS-net and notch filtered datasets, seen in some of the 20 lesions. The input data to the FUS-net and notch filtered datasets occurred during active FUS ‘on’ cycles. Slight tissue relaxation occurring following the onset of FUS ‘off’ could potentially explain sub-micron differences in displacement in the interleaved vs. the FUS-net and notch filtered datasets. Acquisiton of image frames during FUS ‘off’ for the interleaved dataset could have also lessened the presence of cavitation in this dataset, that may have been present in input frames to the FUS-net and notch filtered datasets acquired during FUS ‘on’. This may also contribute to the disparity in CNR between interleaved and FUS-net and notch filtered datasets.

Additionally, a limitation of this study was that both the train and test datasets for the FUS-net model were derived from the same ablation study performed on an ex vivo liver sample (although there was no direct overlap in training and test datasets). The purpose of this study was to serve as a pilot study to study the use of FUS-net for displacement estimation purposes in HMIgFUS thermal ablation. For greater generalizability, a training set with a broader range of ablation parameters and tissue types, and a test set from different acquisitions, can be used.

Conclusions

In this study, we explored three methods (interleaved, notch filter, and FUS-net) of FUS interference filtering from HMIgFUS RF data, using an interleaved AM FUS sequence to generate frames with and without FUS interference. The interleaved method for HMIgFUS was shown to be significantly robust in avoiding FUS interference in all tested cases for FUS ablation monitoring, especially cases with high FUS pressure and long durations, as opposed to traditional notch filtering and FUS-net filtering. CNR obtained from displacement amplitude maps in the interleaved dataset was higher in all cases than that obtained from the notch filtered and FUS-net datasets. Since no further filtering methods or processing methods were applied to this dataset, interleaved HMIgFUS has the potential to be more computationally efficient and can be applied without user input.

FUS-net and notch filtering are advantageous in that they can be used with continuous FUS output, and thus may be more advantageous for low intensity FUS purposes, in which CNR was shown to be higher as compared to high intensity FUS. FUS-net has also shown greater potential in FUS interference filtering while preserving frequency information, as compared to notch filtering. However, in this study, FUS-net was found to produce similar results as notch filtering in displacement estimation, potentially due to ineffective training data or the current architecture of FUS-net, which considers single RF frames at a time, with no information on previous and subsequent frames. Further frameworks can be explored for UNET-based FUS interference filtering for displacement estimation applications.

Data and materials availability

All data associated with this study are available from the corresponding author on reasonable request.