Feature Extraction of Ultrasound Radiofrequency Data for the Classification of the Peripheral Zone of Human Prostate

Abstract

Prostate cancer ranks among the most prevalent types of cancer in males, prompting a demand for early detection and noninvasive diagnostic techniques. This paper explores the potential of ultrasound radiofrequency (RF) data to study different anatomic zones of the prostate. The study leverages RF data’s capacity to capture nuanced acoustic information from clinical transducers. The research focuses on the peripheral zone due to its high susceptibility to cancer. The feasibility of utilizing RF data for classification is evaluated using ex-vivo whole prostate specimens from human patients. Ultrasound data, acquired using a phased array transducer, is processed, and correlated with B-mode images. A range filter is applied to highlight the peripheral zone’s distinct features, observed in both RF data and 3D plots. Radiomic features were extracted from RF data to enhance tissue characterization and segmentation. The study demonstrated RF data’s ability to differentiate tissue structures and emphasizes its potential for prostate tissue classification, addressing the current limitations of ultrasound imaging for prostate management. These findings advocate for the integration of RF data into ultrasound diagnostics, potentially transforming prostate cancer diagnosis and management in the future.

1. INTRODUCTION

Prostate cancer ranks among the most prevalent cancer types affecting men and stands as the second leading cause of cancer-related mortality in American men [1, 2]. The current conventional diagnostic approach involves transrectal ultrasound (TRUS)-guided biopsy, which necessitates the subsequent histopathological analysis of biopsy samples. Unfortunately, this method is both invasive and associated with considerable pain [3]. While noninvasive imaging modalities such as magnetic resonance imaging (MRI), they lack real-time imaging capabilities and tend to be expensive. In contrast, ultrasound (US) imaging is a cost-effective and real-time imaging modality that has been widely used for prostate biopsy and therapy guidance. However, US has limited accuracy for prostate cancer detection.

Recently, studies have demonstrated the utility of raw radiofrequency (RF) data derived from the ultrasound backscatter of clinical transducers for the classification and detection of tumor regions both in vivo and ex vivo [4–7]. Unlike conventional US B-mode images, ultrasound RF signals provide unprocessed and more robust tissue acoustic information, making them a focal point of research in the field of ultrasound tissue characterization. This approach represents a shift towards leveraging the capabilities of RF data to enhance the accuracy and effectiveness of tumor identification through ultrasound imaging. Ultrasound RF data captures the unprocessed acoustic reflections by the ultrasound probe, undergoing a series of signal processing steps before culminating in the creation of the B-mode image. This raw RF data exhibits improved sensitivity to tissue texture, structure, and other pertinent information compared to conventional B-mode images [8–10]. Leveraging RF data proves crucial for the classification and differentiation of various prostate zones.

Research indicates that most prostate cancers originate in the peripheral zone, making it the most common site for prostate cancer development [11, 12]. By segmenting the peripheral zone, valuable information can be extracted for the detection of tumor regions. Examining tissue structure and texture through RF data from the peripheral zone offers insights that contribute to the detection, diagnosis, and assessment of cancer extent. While recent studies have demonstrated the feasibility of prostate zone segmentation using MRI [10], ultrasound encounters challenges in this regard due to limitations in spatial resolution and signal-to-noise ratio.

Recent studies have affirmed the feasibility of incorporating radiomics into ultrasound imaging with noteworthy findings [13–15]. Radiomics stands out as a robust tool for extracting features from images, encompassing elements like texture, pixel intensities, and other critical attributes [16, 17].

In this study, we show the feasibility of using ultrasound RF data for segmenting different regions of ex-vivo prostate specimens, specifically the peripheral zone, using radiomics and image filters. The preliminary results show that the peripheral zone of the ex-vivo prostate can be identified and differentiated from the rest of the prostate using radiomics, 3D surface plots and some texture filters. We have shown the preliminary results from the radiomics feature maps, which demonstrate the distinction between the two zones of the prostates [18].

2. MATERIALS AND METHODS

2.1. Ultrasound data acquisition of ex-vivo prostate specimens

Ultrasound data of the whole prostate specimens were collected using the Verasonics Vantage System 256 (Verasonics, Redmond, WA) and GEM5ScD phased array transducer. We have acquired ultrasound data of a total of 4 ex vivo prostates of human patients. The prostate samples were submerged in a water bath as shown in Figure 1. The transducer was attached to an automated movable holder and placed on top of the prostate. We collected three-dimensional (3D) ultrasound data of the prostate by moving the transducer from one end to another with a fixed step size while the frames were being captured. Both B-mode and the raw RF data were captured and saved for each sample.

Figure 1.

Prostate specimen submerged in a water bath.

2.2. RF data and ultrasound image processing

We used a wide beam imaging script for the data acquisition for better imaging quality. The RF data that were saved using the wide beam imaging script had the data of multiple beams at various angles. We chose the average of all beams with zero-degree angles. The acquired B-mode images of the prostate were processed to improve the quality of the image. The images were cropped to the prostate area as the region of interest. The images were also background subtracted to remove noise using averaging all images acquired at the same location. The contrast and brightness were increased to highlight the prostate in the B-mode image.

2.3. Correlation of B-mode images and RF data

B-mode images were utilized to segment the center and peripheral zones of the prostate, facilitating radiomic feature extraction. We used B-mode images instead of the RF data as it is challenging for humans to interpret. Both RF data and B-mode images were captured from Verasonics. Following correlation with B-mode data, the RF data was cropped to the region of interest. MATLAB was employed to visually assess and extract masks, ensuring correlation between RF Data and B-mode images. The correlated B-mode and RF data are depicted in Figure 2. We used annotated histology slides to identify the peripheral zone of the prostate samples. One of the middle slices from the histology slides was correlated with the B-mode image and the RF data frame.

Figure 2.

B-mode image and RF data after correlation for visual comparison.

2.4. Feature extraction of prostate peripheral zone using RF data.

The spatial information from B-mode images within the same correlated RF data and B-mode images was utilized to selectively crop the RF data to the region of interest. Subsequently, the cropped RF data was graphically represented in the axial plane, enhancing the visualization of the prostate’s peripheral zone, as illustrated in Figure 2. By employing a range filter in MATLAB, we extracted pertinent features to accentuate the peripheral zone of the prostate. The range filter calculates the range of the pixel intensities in the surrounding which enhances the edges. Additionally, a 3D plot was employed to visualize the pixel intensities of the RF data frames. Comparative analysis was conducted between the regions corresponding to the peripheral zone of the prostate and the remainder of the prostate.

2.5. Radiomics feature extraction and comparison.

We employed radiomic feature extraction on the RF data and captured various texture, shape, and statistical characteristics to characterize the underlying tissue properties. We used two separate regions of interest, namely the peripheral zone and the central zone, to compare the features for the distinction between the two zones. Radiomic features were extracted from the delineated regions using a comprehensive feature extraction process. We used pyradiomics for the feature extraction [18]. The extracted features included shape descriptors such as elongation, sphericity, and surface area. Texture features were derived through gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM), and gray level size zone matrix (GLSZM) analyses [18]. Additionally, first-order statistics and gray level dependence matrix (GLDM) features were calculated to capture intensity distribution and dependence relationships within the regions.

3. RESULTS

3.1. Image filters applied on RF data to highlight the peripheral zone.

Figure 3 displays the RF data image after the application of a range filter. Notably, the pixels in the peripheral zone exhibit an expanded dynamic range compared to the surrounding prostate area. This extended dynamic range of peripheral zone pixels signifies a greater capacity to capture comprehensive acoustic details. Furthermore, certain segments of the peripheral zone exhibited relatively higher brightness, indicating captured alterations in tissue composition. Figure 3 highlights a clear difference in the RF data from the peripheral zone when compared to the rest of the prostate. This emphasizes that there are notable variations in the structure of the tissues.

Figure 3.

RF data processed by range filter and with highlighted peripheral zone.

3.2. Three dimensional plots of the RF data

The 3D graphs presented in Figure 4 visually demonstrate that the peripheral zone exhibits a greater dynamic range, aligning with the RF data after the range filter processing. Furthermore, the graph highlights the noticeable characteristics of the peripheral zone in contrast to the adjacent prostate tissue. It is evident that the pixel intensities within the peripheral zone show greater variability compared to the surrounding prostate regions. Notably, the dynamic range of pixels in the peripheral zone is comparatively elevated. It is also evident that the region of the peripheral zone can be differentiated from the rest of the prostate using the RF data with filters and 3D plots.

Figure 4.

3D plots of RF data with highlighted peripheral zone.

3.3. Comparison radiomic features of two zones.

We present an initial comparative visualization of radiomic features extracted from two distinct regions of interest, as depicted in Table 1. The radiomics features analysis was conducted using four specimens. Masks were employed to crop the peripheral zone and central zone, and radiomic features for both zones were generated using pyradiomics [18–21]. Table 1 illustrates the comparison of peripheral zone and central zone for all four specimens, utilizing heatmaps to represent features, where red corresponds to high values and green to low values. Notable differences were observed in features between peripheral zone and central zone, particularly in original first-order energy and original first-order total energy, showing consistent significant disparities across all four specimens. Here, energy serves as a measure of voxel value magnitudes in an image [21]. Larger values indicate a greater sum of the squares of these values. For Specimens 1, 3 and 4, first-order energy was higher in the peripheral zone, while for Specimen 2, the central zone exhibited a higher value. Similarly, the total energy of the peripheral zone was higher in Specimens 1, 3 and 4, and lower for Specimen 2. Original GLDM features consistently displayed higher values in the peripheral zone for Specimens 1, 3 and 4, except for original GLDM Large Dependence Emphasis, where the central zone consistently exhibited higher values. GLRLM features also exhibited significant differences between the two zones. The radiomic features showcased in our analysis emphasize significant differences between the peripheral zone and central zone of the prostate. These distinctions offer a solid foundation for effectively distinguishing between these two zones. The identified radiomic features serve as robust indicators, contributing to a reliable basis for zone differentiation.

Table 1.

Radiomics feature heatmaps of peripheral zone and central zone of the prostate of four specimens.

4. DISCUSSION AND CONCLUSION

In this study, we showed that RF data from ultrasound imaging can be successfully used to differentiate the peripheral zone from the rest of the prostate. We used a range filter to highlight the peripheral zone and used 3D plots to show the dynamic range of pixel intensities in the peripheral zone, which were absent in the other regions of the prostate. We also proposed a radiomics approach to extract the features from the RF data of the prostate. Radiomics gives us much more differentiational information at different zones of the prostate from the RF data. It extracts a multitude of quantitative features from the images, such as texture, shape, and intensity patterns. Radiomics also offers the potential to enhance segmentation of all the four zones of the prostate owing to different tissue structure and contributing to a more comprehensive understanding of underlying tissue properties. We particularly saw differences between the peripheral central zones in the first order, GLDM, GLRLM, GLSZM, NGTDM features. Each of these features provides complementary information from the pixel intensities of the region of interest.

The preliminary results presented in this study demonstrate the ability of RF data to detect alterations in tissue structure and differentiate the peripheral zone from the remainder of the prostate. The application of a basic image filter reveals the potential to highlight the peripheral zone. These results imply that RF data possesses the capability to capture more comprehensive acoustic information compared to beamformed data. This is attributed to the utilization of unprocessed raw signals, which constitute the authentic acoustic information scattered back from the tissue. The increased acoustic information holds the potential for more precise tissue characterization and potentially improved accuracy in the identification and diagnosis of prostate tumor regions. The adoption of RF data for ultrasound imaging diagnosis also holds promise for broad applications.