Hemoglobin Microbubbles and the Prediction of Different Oxygen Levels Using RF Data and Deep Learning

Abstract

Ultrasound contrast agents (UCA) are gas-encapsulated microspheres that oscillate volumetrically when exposed to an ultrasound field producing backscattered signals efficiently, which can be used for improved ultrasound imaging and drug delivery applications. We developed a novel oxygen-sensitive hemoglobin-shell microbubble designed to acoustically detect blood oxygen levels. We hypothesize that structural change in hemoglobin caused due to varying oxygen levels in the body can lead to mechanical changes in the shell of the UCA. This can produce detectable changes in the acoustic response that can be used for measuring oxygen levels in the body. In this study, we have shown that oxygenated hemoglobin microbubbles can be differentiated from deoxygenated hemoglobin microbubbles using a 1D convolutional neural network using radiofrequency (RF) data. We were able to classify RF data from oxygenated and deoxygenated hemoglobin microbubbles into the two classes with a testing accuracy of 90.15%. The results suggest that oxygen content in hemoglobin affects the acoustical response and may be used for determining oxygen levels and thus could open many applications, including evaluating hypoxic regions in tumors and the brain, among other blood-oxygen-level-dependent imaging applications.

1. INTRODUCTION

Ultrasound contrast agents (UCAs), also known as “microbubbles” are used in contrast-enhanced ultrasound imaging due to their unique scattering properties [1]. UCAs are vascular contrast agents which enhance the scattered signals in the blood in ultrasound images. When UCAs are exposed to an ultrasound field, they produce non-linear acoustic response that can be detected using a clinical transducer. Due to their unique acoustic response, they are used for enhancing ultrasound images [3]. UCAs are used for cancer diagnosis and other biomedical applications [2–6]. Contrast enhanced ultrasound imaging (CEUS) is one such application, which is established for cancer diagnosis [6]. CEUS imaging is also used for biosensing applications. As its cost effectiveness and the availability of ultrasound scanners, it has become one of the desirable modalities for biosensing applications. CEUS imaging depends on UCAs to detect changes in blood volume and the direction of blood flow. Most of these UCAs are designed to generate backscatters that can be detected by an ultrasound probe to determine the blood characteristics like the direction, flowrate, and volume but not the percentage of oxygen present or the hypoxic regions. In vivo biosensing of oxygen level is important to monitor general health and patient response to therapy [7]. There is a need to develop novel UCAs that can act as a biosensor to detect oxygen levels.

Other imaging methods, such as magnetic resonance imaging (MRI) [8–10], positron emission topography (PET) [11–15], and optical imaging [16], have been explored to measure oxygen levels or hypoxia tissue. For MRI and PET, the cost can be high and real-time imaging can be difficult [17–18]. PET imaging also has a disadvantage of radiation exposure [18]. Biosensors with acoustical properties that are detected using ultrasound can be attractive because of its low cost, real time, no radiation. Compared to optical approaches, ultrasound imaging is capable of scan deep tissues in vivo [19].

Recently, our group developed and optimized a novel oxygen-sensitive hemoglobin-shell microbubble (HbMBs) designed to acoustically detect blood oxygen levels in vivo. Using high-intensity ultrasound, Wong and Suslick reported a method for the synthesis of air-filled hemoglobin (Hb) microbubbles [20]. We hypothesized that the structural changes in the hemoglobin in response to varying oxygen levels will alter the mechanical properties of the bubble shell, resulting in detectable changes in the bubble acoustic signature. Specifically, varying oxygen levels can cause the mechanical changes in the hemoglobin shell, which can cause the microbubble to oscillate distinctly when exposed to an ultrasound field. The oscillation may produce an acoustical difference that can be detectable using an ultrasound transducer. Therefore, we can use this acoustic property to measure the oxygen levels in vivo using ultrasound imaging.

In this study, we focused on the development of a deep learning algorithm to distinguish between the acoustic responses of hemoglobin microbubbles with different oxygen levels using the raw one-dimensional radiofrequency (RF) data that were acquired using a clinical transducer. The goal was to demonstrate that changes in oxygen levels in the solution can be detected using hemoglobin microbubbles. To the best of our knowledge, this is the first study on deep learning for differentiating oxygen levels using hemoglobin microbubbles.

2. METHODS

2.1. Preparation of Hemoglobin Microbubbles

Hemoglobin stock solution is prepared with 10mg/mL of hemoglobin powder dissolved in 10% glycerol 10% polypropylene glycol PBS solution. Tryptophan stock solution is prepared separately following the same procedure mentioned above. The stock solutions are homogenized and mixed at a ratio of 4:1 (Hemoglobin: Tryptophan). This solution is heated until it reaches 50°C. After the solution reaches 50°C, we sonicate it for 5 seconds and cool it immediately by immersing the solution in an ice bath. After the solution cools down completely, it is reduced by adding Sodium dithionite, Sodium sulfate stock solution and then it is oxygen saturated for 3 minutes by running oxygen gas on the surface of the solution. At this point, the solution turns blood red. We prepared and used a total of eight samples for this experiment.

After the HbMBs are prepared, separate solutions are made for the oxygenated hemoglobin MBs (Oxy MBs) and the deoxygenated hemoglobin MBs (Deoxy MBs). Oxygenated sample is prepared by saturating PBS with oxygen for 3 minutes and then HbMBs are mixed at a fixed concentration of 2.5×10⁶ MBs/mL. Similarly, the deoxygenated sample is prepared by saturating PBS with nitrogen for 30 minutes until the oxygen percentage reaches below the threshold (10%) and is mixed with HbMBs at a fixed concentration of 2.5×10⁶ MBs/mL. Both oxygenated and deoxygenated samples are transferred to a 50 mL syringe with care.

2.2. Data collection Using Verasonics Vantage

After preparing oxygenated and deoxygenated solutions containing HbMBs, the raw RF data was collected using a GEM5ScD clinical phased array transducer and a Verasonics Vantage 256 system (Verasonics, Redmond, WA). The concentration of UCAs and the acoustical pressure for both samples are kept constant to reduce external variations while collecting the RF data. Figure 1A shows the flow phantom setup that is used. The phantom has two tunnels that run parallel to each other. Both the tunnels carry Oxygenated and Deoxygenated MBs solution respectively and they are exposed to ultrasound at the same time and the RF data of both tunnels are captured together in a single frame. The complete process of collecting data for Oxygenated and Deoxygenated MBs is completed simultaneously. This reduces the variability in data that might occur while collecting datasets subsequently. A 50mL syringe is used to pump the solution through the phantom and was attached to a syringe pump set at a 2mL/min constant flow rate. The tunnels carrying the samples were swapped after every experiment. This reduced any variations that might be caused due to the tunnels.

Figure 1:

1A: Flow phantom setup. 1B: 2D image from the Verasonics system with ROI highlighted. 1C: Corresponding 1D RF data plot of a single element central to the tunnel for the oxygenated hemoglobin microbubbles (Oxy HBMb) and the deoxygenated hemoglobin microbubbles (DeOxy HBMb).

The oxygen percentage was measured at the end of each data acquisition for both oxygenated and deoxygenated samples. We used a dissolved oxygen (DO) meter to calculate the oxygen levels in the microbubble solution. The DO meter was calibrated to the atmospheric oxygen level at 100%. The oxygen-saturated PBS measured more than 200% and the deoxygenated PBS measured less than 10% before the start of the experiment. The oxygen percentages were also measured after the data acquisition to make sure that the HbMBs were in oxygenated and deoxygenated environment during the data acquisition as shown in Figure 2, black bars represent the oxygen percentage for deoxygenated HbMBs and white represents the oxygen percentage for oxygenated HbMBs. The oxygen-saturated samples had oxygen levels of around 150%. The nitrogen-saturated samples had oxygen levels of around 60%.

Figure 2:

Oxygen percentage with respect to the atmospheric oxygen levels (100%) for the oxygenated and deoxygenated samples.

2.3. Data Preprocessing

The steps and conditions followed for data preprocessing were the same for all the samples. The captured RF data went through a series of data preprocessing methods.

2.3.1. Training and Validation Dataset

The data processing technique followed here is similar to the method, as we reported earlier [22]. The MAT files were acquired using the Verasonics Vantage system 256 and each file had 10,000 frames in one file and there were 13 files in total that were acquired with multiple HbMBs samples. These files were the raw RF data that were saved into a folder. After acquiring and saving the RF data, a single element that is the central element for the tunnel carrying Oxygenated MBs and Deoxygenated MBs was saved into separate files. After the separation of RF data into oxygenated and deoxygenated, the RF data was cropped to the region of interest (ROI) as shown in Figure 1B to avoid the walls of the tunnels that might affect the results and then they were loaded in Python and converted into a data frame. In total there were 260,000 amplitude responses (frames) of Oxygenated and 260,000 amplitude responses (frames) of Deoxygenated respectively. These datasets were separated into different classes and labeled as class A and class B. After labeling the data, they were split into training and validation datasets with 92:8.

2.3.2. Test Dataset

The testing dataset for this study was acquired separately on a fresh set of samples and on a different day. The testing dataset was not used in the training or validation of the deep learning model. The samples were prepared with a fresh batch of HbMBs and oxygenated and deoxygenated PBS. This process was followed to check for the efficiency of the 1D CNN model. The preprocessing steps were followed according to the same protocol that was followed for the training datasets. The cropped RF datasets were separated into oxygenated and deoxygenated groups and followed by normalization as there were two separate samples used for the test data acquisition. A total of 2000 frames were used for testing the model. After normalizing the data, the two classes were labeled and separated, and the model was evaluated for performance.

2.4. 1D Convolutional Neural Network Architecture

A 1D convolutional neural network was used for the classification task of oxygenated and deoxygenated HbMBs [21]. The model consists of four 1D convolution layers, one max pooling layer, and a fully connected layer. Figure 3 shows a detailed description of the 1D CNN network architecture used. After every convolution layer, we used batch normalization to avoid overfitting and improve model performance. GELU activation function [24] was used for all the convolution layers and binary cross entropy loss function with Adam optimizer was used.

Figure 3:

Detailed 1D CNN Network architecture with 1D convolution, max pooling, and flattening layers. 1D RF data is given as input and the frames are stacked on top of each other as shown.

The task here was to classify oxygenated MBs and Deoxygenated MBs with high accuracy. Classifying them into separate classes proves that the acoustic response is distinguishable for Oxygenated and Deoxygenated MBs, and they respond differently when exposed to an ultrasound field. The labeled dataset after splitting it into training and validation datasets was fed into the algorithm, with each sample point in the frame as an input. The output layer had one unit (two classes) with a GELU activation function. We implemented the 1D CNN in Keras and TensorFlow [25].

The data for the input of the 1D CNN model is shown in Figure 3. The input is the 1D RF data that is cropped for the region of interest. Each frame has a total of 138 data points that correspond to the amplitude values along the depth. These amplitude values are the RF data from a single element of the transducer that runs central to the tunnel and has been cropped. These frames are then stacked on top of each other giving us a matrix of 260,000 by 138. These frames are provided as the input to the 1D CNN model, where a single frame is given as an input.

3. RESULTS

3.1. Training and Validation Results

A total of 260,000 frames were used for training and validation of the 1D CNN model. The datasets were split into training and validation datasets with 92:8. A total of 50 epochs were used to train the model. We achieved a training and validation accuracy of 98.72% and 86.8% with this model. The sensitivity and specificity of classification were 95.3% and 92.8% respectively. The results suggest that the acoustic response of oxygenated and deoxygenated HbMBs to ultrasound can be classified and distinguished between each other, showing that the change in oxygen percentage in the surrounding for HbMBs does impact the structural changes in the shell and produce a distinguishable acoustic response that can be detected using a clinical transducer.

The training and validation datasets were collected on different days using independently generated samples. The microbubble size distribution was different for all the samples which will impart variations in the acoustic response. The dataset was normalized in the preprocessing step to account for these variations. The tunnels carrying the microbubbles were swapped after every sample. The model achieved a training accuracy of 98.72% at the end of the 50^th epoch. The validation accuracy was 86.8% at the end of the 50^th epoch. The model reached above 90% accuracy after 4 epochs and gradually reached 98%. The loss used for the model was binary cross entropy. The model quickly optimizes the loss and falls below 0.1 at the end of the 4th epoch. The model converges the loss for training and validation below 0.05 by the 15^th epoch. We used an adaptive learning rate to optimize the training of the model.

3.2. Testing Results

The testing data was separated from the training and validation data. A total of 2000 frames of oxygenated and deoxygenated respectively were used for evaluating the performance of the deep learning model. An accuracy of 90.15% was achieved with this testing dataset. The sensitivity and specificity are 92.47% and 87.83% respectively for the model. The dataset used here was separate from the training and validation and was used only in the testing and evaluation of the model.

The results show that after the deep learning model was trained, it can classify the acoustic response of oxygenated and deoxygenated HbMBs. The different oxygen levels in the surrounding may have affected the shell of HbMBs, which cause structural changes and lead to distinct acoustic signals that were captured by the transducer.

4. DISCUSSION AND CONCLUSION

The study demonstrated that the acoustic response of hemoglobin microbubbles with two oxygen levels is distinguishable using our deep learning model. In this study, we used oxygen levels at the extreme conditions: Oxygenated MBs were oxygen-saturated; and deoxygenated MBs were nitrogen saturated. Our CNN model was able to distinguish the acoustic response between oxygenated and deoxygenated hemoglobin microbubbles. With a bigger dataset, the validation accuracy can be improved. The variations in the HbMBs size distribution may result in variations in the acoustic response.

In our future work, we plan to develop an algorithm that can quantify oxygen levels based on the acoustic response of the hemoglobin MBs. This approach can be potentially used to detect oxygen levels in vivo by acquiring the acoustic response of hemoglobin MBs in the blood pool, which may provide information on oxygen levels in the body. The hemoglobin MBs and deep learning method can be applied to evaluate hypoxic tissues in tumors and in the brain, thereby opening numerous medical applications in the future.

In conclusion, we developed a deep learning model and a novel hemoglobin MBs, which can be used to distinguish different oxygen levels. Though the model is trained for the extreme conditions of oxygen levels, the study suggests that deep learning methods may be able to predict different oxygen levels by using more datasets and advanced neural networks.