H-scan discrimination for tumor microenvironmental heterogeneity in melanoma

Abstract

Objective:

Melanoma is a form of malignant skin cancer that demonstrates significant intertumoral differences in the tumor microenvironments (TME) secondary to genetic mutations. The heterogeneity may be subtle but can complicate the treatment of metastatic melanoma, contributing to a high mortality rate. Therefore, developing an accurate and non-invasive procedure to discriminate microenvironmental heterogeneity to facilitate therapy selection is an important goal.

Methods:

In vivo murine melanoma models that recapitulate human disease using synchronous implanted YUMM 1.7 (Yale University Mouse Melanoma) and YUMMER 1.7 (Yale University Mouse Melanoma Exposed to Radiation) murine melanoma lines were investigated. Mice were treated with antibodies to modulate the immune response and longitudinally scanned by ultrasound (US). US radiofrequency data were processed for the H-scan analysis, attenuation estimation, and B-mode processing to extract five US features. The measures were used to compare different TMEs (YUMMER vs. YUMM) and responses to immunomodulatory therapies with CD8 depletion or PD-1 (programmed cell death protein-1) inhibition.

Results:

Multiparametric analysis produced a combined H-scan parameter, resolving significant differences (1) between untreated YUMMER and YUMM, and (2) between untreated, PD1-treated, and CD8-treated YUMMER. However, more importantly, the B-mode and attenuation measures failed to differentiate YUMMER and YUMM and to monitor treatment responses, indicating that H-scan is required to differentiate subtle differences within the TME.

Conclusion:

We anticipate that the H-scan analysis could discriminate heterogeneous melanoma metastases and guide diagnosis and treatment selection, potentially reducing the need for invasive biopsies or immunological procedures.

INTRODUCTION

Medical ultrasound (US) has been widely used to diagnose and screen diseases, including liver steatosis/fibrosis/tumors, kidney fibrosis, and breast cancer. Usage of ultrasound has increased due to its numerous advantages: it is non-invasive, it provides real-time imaging, it can be portable, and it has a lower cost compared with magnetic resonance (MR) and computed tomography (CT). However, US traditionally has limited usage in disease diagnosis because its qualitative imaging display relies on clinicians’ interpretation. Further, traditional B-mode images alone cannot detect subtle pathological changes in tissues. B-mode imaging also has lower diagnostic performance than MR or CT [1], which motivates the development of methodologies that extract more information from raw US signals. Recently, quantitative US (QUS) has been introduced to provide quantitative measures from US scans [2], which enabled diagnostic improvement compared to traditional B-mode imaging in many applications such as cancer and disease detection, disease progression monitoring, and treatment response tracking [3–5]. The diagnostic accuracy of QUS can be verified using accepted measures from pathology or MR. For steatosis, MR imaging-driven proton density fat fraction (MRI-PDFF) is known as one of the accurate approaches among non-invasive imaging modalities [6–10]. Thus, QUS parameters extracted from steatosis human subjects were compared with MRI-PDFF and demonstrated high correlation [5]. Although QUS studies includes spectral based analysis, another US frequency-dependent analysis of the H-scan has been proposed [11, 12]. The H-scan uses a matched filter analysis to characterize scattering behavior and provides color-coded images to visualize different categories of scatterers. The H-scan has been applied to liver imaging to monitor pancreatic cancer metastasis [13]. That study demonstrated that the H-scan results showed a strong agreement with biofluorescence measurements and better performance compared to US measurements of shear wave speed. Moreover, it has been revealed that the H-scan measures have high correlations with pathology or MRI-PDFF for assessing many diseases, including liver steatosis (correlation coefficient, R = 0.83, with histology fat fraction [14] and R = 0.86 with MRI-PDFF [15]), kidney fibrosis (R = 0.99 with histology fibrosis score [16]), and breast cancer (Area under the curve (AUC) = 0.81 with biopsy). These suggested the potential of H-scan to contribute diagnosis based on US quantitative measurements.

As a great number of quantitative parameters have been introduced to identify tissue characteristics, the integration of multiple parameters may improve characterization accuracy. Thus, multiparametric analysis was suggested with several approaches: (1) non-linear transformation to calculate a combined parameter [17], (2) support vector machine (SVM) to incorporate information from parameters [18, 19], and (3) principal component analysis and inner product to obtain a 1D combined parameter [14]. These studies demonstrated that incorporating multiple parameters improved differentiation between different stages of disease, for example: (1) normal vs. inflammation vs. fibrosis in the liver [18]; (2) normal vs steatosis in the liver [20]; (3) benign vs. malignant breast lesion [21, 22]. Moreover, these studies showed that the H-scan was one of the most accurate contributors to tissue classification.

Although previous studies investigated different disease conditions, the efficacy of US to discriminate subtypes of the same disease remained unknown. Intertumoral heterogeneity in metastatic melanoma is a clinically relevant phenomenon that requires such distinction. Metastatic melanoma patients have an average survival rate of less than six months [23]. One contributor to mortality rate may be heterogenous treatment responses and different metastasis exhibited within the same individual secondary to genetic and tumor microenvironmental (TME) variations between the two lesions [24, 25].

In order to study tumor heterogeneity in melanoma, murine models of melanoma were previously developed: the Yale University Mouse Melanoma line, known as YUMM 1.7 (YUMM) and the Yale University Mouse Melanoma Exposed to Radiation, also known as YUMMER 1.7 (YUMMER). These cell lines are used to recapitulate the genetic and TME heterogeneity [25]. The YUMM and YUMMER cell lines share 37% of synchronous mutations [26], which created immunologically distinct TMEs, with YUMMER having increased immune infiltration compared to YUMM [25]. Detection of the distinct subtypes requires an invasive biopsy for comprehensive analyses such as immunohistochemistry, Luminex analyte assay, and RNA sequencing; these biopsies are typically clinically cost-inefficient. However, given the effect of immunologically distinct TME on melanoma therapy response, non-invasive measures are needed to identify immunological properties of metastases to improve therapy selection.

In this study, we used non-invasive H-scan US approaches to detect TME heterogeneity in a murine model of melanoma with implanted YUMM or YUMMER tumors. To study the efficiency of H-scan on differentiating immunological differences in the TME, YUMM and YUMMER tumors were treated with either anti-CD8 antibodies, which decrease the amount of intratumoral immune infiltration, or with anti-PD-1 (programmed cell death protein 1) antibodies, which increases immune infiltration. The H-scan analysis was performed to discriminate YUMM and YUMMER and to monitor treatment responses over time. H-scan parameters and US parameters were measured, and a multiparametric analysis was performed to combine parameter information.

MATERIALS AND METHODS

Study Design

This study protocol was approved by the University of Rochester’s University Committee on Animal Resources. The bilateral flanks of six- to eight-week-old female wildtype C57BL/6J mice from the Jackson Laboratory (Bar Harbor, ME, USA) were injected with one million YUMM 1.7 or YUMMER 1.7 cells [25]. The implanted murine melanoma tumors were ultrasound-scanned, as shown in Fig. 1. Mice were split into isotype antibody treated control, anti-PD-1 treated, and CD8-depleted groups to represent baseline immune infiltration, increased immune infiltration, and decreased immune infiltration, respectively. The mice received treatment three-times per week once the tumors reached 4 mm in the largest diameter and were sacrificed on day 27 post-tumor implantation; all tumors had a largest diameter of between 4 mm and 2 mm at the time of scanning. Tumors subsequently underwent hematoxylin and eosin (H&E) histological staining and CD45 immunohistochemical staining.

Figure 1.

Murine melanoma model study design and experimental setup. (a) Tumor injection in mice flanks, tumor growth, and ultrasound scan. Illustrations of (b) an anesthetized mouse with two induced tumors on both sides of flanks before ultrasound scan and (c) scanning a grown tumor with ultrasound gel. (d) 3D scan was performed with a mechanical motion controller.

US Acquisitions

Ultrasound scanning was performed using the Vevo 3100 imaging system (FUJIFILM VisualSonics, Inc., Toronto, Canada) equipped with a 40 MHz center frequency linear transducer (MX550D). Focused beam transmission was used, and a single focal depth of 10 mm was set for all scans. We acquired 3D volume images with a 0.05 mm step size between 2D frames, and the acquired volumes had approximately 0.1 mm to 5 mm lengths depending on tumor size, whereby each 3D volume data point had approximately 2 to 100 frames; the 3D scan with the linear transducer was performed with a mechanical motion controller as shown in Fig 1 (d). The machine saved radiofrequency (RF) data, and we produced B-mode images by in-phase and quadrature (IQ) demodulation, envelope detection, and log compression. The B-mode images were used for manual contouring of the melanoma boundary for each frame, which was the region of interest (ROI) for this study, indicating the melanoma area. The saved RF data format was utilized for H-scan processing and attenuation correction, and envelope and log-compressed data were used to measure intensity-based parameters.

We estimated two B-scan parameters: B-scan intensity and B-scan signal-to-noise ratio (SNR). First, B-scan intensity was obtained by averaging the dB scale intensity of log-compressed data within the contoured ROI. B-scan SNR was calculated utilizing the envelope data:

$$B-scanSNR=\frac{{\mu }_B}{{\sigma }_B}$$ (1)

where ${\mu }_B$ and ${\sigma }_B$ are the average and standard deviation of the envelope data within an ROI, respectively. When estimating the two B-scan parameters, as illustrated in Fig 2, we excluded any area having ultrasound artifacts, such as reverberation artifacts, shadowing, and non-uniform US beam energy in the depth direction due to focused transmission. Our scan setting with a single focus at 10 mm reduced the SNR at deeper depths compared to the focal point, and thus our parameter estimation occurred at depths less than 11 mm. All scanlines and samples having reverberation artifacts or shadowing were excluded from the estimation by defining lateral and axial intensity thresholds, as depicted in Fig 2. Fig 2 also displays a mask example and US B-scan speckle within the mask.

Figure 2.

Exclusion of hypoechoic (low signal) area to estimate B-scan parameters.

H-scan analysis

We extracted features from the RF data frequency using the H-scan analysis [12]. The H-scan was performed to investigate frequency spectral shifts caused by attenuation or scatterer size changes. The process is summarized in Fig 3. First, US RF data were acquired and used as input for the H-scan. Because US propagation causes frequency downshifts along depth, leading to confusion between the attenuation effects and scattering signature changes we aim to detect, attenuation correction was performed. The frequency-dependent attenuation can be modeled by $e^{-\alpha fx}$ where $\alpha$ is attenuation coefficient in Np/cm, $f$ is frequency in MHz, and $x$ is depth in cm. Conceptually, by multiplying $e^{+\alpha fx}$ to the frequency spectrum $S\text{(}f)=fft\text{(}RF\text{(}x\text{))}$, we can obtain attenuation-corrected (ac) RF data $\left(R{F}^{ac}\right)$:

$$R{F}^{ac}=ifft\left(S(f)\cdot e^{+afx}\right).$$ (2)

However, equation (2) contains two variables, frequency $f$ and depth $x$. In order to make the depth variable a constant, we divided the depth ROI into 10 zones, and each zone $z$ had a representative depth $x_z$. Now, a simple calculation of attenuation-corrected RF data is possible with $S(f)\cdot e^{+\alpha f{x}_z}$, and each zone’s attenuation-corrected RF data is $R{F}_z^{ac}(t)=S(f)\cdot e^{+\alpha f{x}_z}$. To obtain the attenuation-corrected RF data $R{F}^{ac}(t)$ over the entire depth range/zones, $R{F}_z^{ac}(t)$ can be combined. The attenuation equation $e^{-\alpha fx}$ also describes intensity attenuation, although most US machines provide RF data after time-gain-compensation (TGC), so $S(f)\cdot e^{+\alpha fx}$ compensates for intensity attenuation. Thus, before combining, intensity normalization between zones is needed only if the RF data was acquired after TGC. The H-scan processing used attenuation-corrected RF data: $R{F}^{ac}(t)$. For matched filtering, 256 Gaussian filters ($G_i$ for $i=1,2,\dots ,256$) were determined with different peak frequencies: $f_{p\_i}$ for $i=1,2,\dots ,256$ where $f_{p_{-}1}<f_{p_{-}2}<\cdots <f_{p_{-}256}$ and $f_{p_{-}k+1}-f_{p_{-}k}=\Delta f\cdot \Delta f$ is a constant which can be defined based on the frequency spectrum of the RF data for all peak frequencies to cover the entire spectrum from low to high frequency components. A bandwidth of the Gaussian filters was specified by a standard deviation, ${\sigma }_G$, as detailed in Fig. 3. The bandwidth of Gaussian filters was set to have 70% of the spectral bandwidth; we first estimated an averaged standard deviation of frequency spectra for RF data $\left({\sigma }_{RF}\right)$ to find the bandwidth, and the bandwidth of the Gaussian filters were determined by ${\sigma }_G={\sigma }_{RF}\cdot 0.7$. Bandpass filtering between the Fourier transform of RF data and the Gaussian filters was performed, and the 256 outputs were inverse Fourier-transformed, resulting in 256 outputs of the matched filtering: $M{F}_i$ for $i=1,2,\dots ,256$. Note that convolution in the time domain can be used instead of bandpass filtering in the frequency domain. After the filtering, each time sample $t$ in the axial direction (each pixel in 2D US images) had 256 values: $M{F}_1(t),M{F}_2(t),\dots ,M{F}_{256}(t)$. We can find a unique maximum of $M{F}_{i_{max}}(t)$ where $i_{max}$ is the index of the Gaussian filter $G_{i_{max}}$, and $G_{i_{max}}$ has a peak frequency $f_{p_{-}{i}_{max}}$ as shown in Fig 3. Then $f_{p_{-}{i}_{max}}$ is the estimated frequency component at time sample $t$: $f_p(t)=f_{p_{-}{i}_{max}}$. The estimated frequency components were mapped into H-scan color levels using the color bar shown in the pseudo color mapping block in Fig 3. The H-scan color levels from 1 to 256 are depicted as more red to more blue, in order, as shown in the color bar in Fig 3. The redder color indicates low frequency components and larger scatterers, corresponding to the lower color levels. Thus, the H-scan estimates a color level for each pixel within a ROI. Since the color levels range from 1 to 256, the color levels from 1 to 128 were classified as red pixels; those from 129 to 256 were classified as blue pixels. We defined H-scan blue percentages (H-scan % blue) as:

$$H-scan\%blue=\frac{numberofbluepixels}{totalnumberofpixelswithinROI}\times 100\%.$$ (3)

Figure 3.

H-scan processing.

We calculated H-scan SNR as:

$$H-scanSNR=\frac{{\mu }_H}{{\sigma }_H}$$ (4)

where ${\mu }_H$ and ${\sigma }_H$ are the average and standard deviation of the H-scan color levels within a ROI, respectively.

As previously mentioned, when using raw RF data as the H-scan input, the attenuation effect of the frequency downshift can be detected, showing more red color for deeper depths. We estimated attenuation coefficient using the H-scan blocks in Fig. 3. Attenuation estimation used raw RF data before attenuation correction. Through the matched filtering and peak frequency detection blocks, peak frequency $\left(f_p(x)\right)$ along depth $(x)$ was estimated, and attenuation coefficient $(\widehat{\alpha }(x)[\text{dB}/\text{MHz}/\text{cm}])$ can be calculated by:

$$\widehat{\alpha }(x)=-\frac{f_p(x)-f_0}{x\cdot {\sigma }^2}$$ (5)

where $f_0$ is center frequency and $\sigma$ is band width of transmit ultrasound beam. More details for attenuation estimation, including equation (5), are found in [19]. In this study, we first estimated the attenuation coefficient, and the estimated attenuation coefficient was used for attenuation correction.

Multiparametric analysis

For the multiparametric analysis, we estimated five parameters: two B-scan parameters (B-scan intensity and B-scan SNR) based on US backscattering, two H-scan parameters (H-scan % blue and H-scan SNR) extracted from US frequency-dependent information, and the attenuation coefficient reflected by US physics. By performing multiparametric analysis, the best performing parameters were selected and combined; detailed procedures are provided in Fig 4.

Figure 4.

Multiparametric analysis to combine information from parameters.

As shown in Fig 4, the extracted raw parameters had different scales. Z-score normalization was performed to obtain 0 mean and 1 standard deviation, which enabled a reasonable comparison between parameters with a consistent scale and distribution. Feature selection was performed by investigating possible parameter combinations, such as all five parameters, four parameters after excluding only one, three parameters after excluding two, and two parameters. For each of the possible combinations, principal component analysis (PCA) calculated the first and second principal components (PC1 and PC2, respectively). PC1 is considered as a combined parameter for statistical analysis. One-way ANOVA was used to evaluate differentiation between (1) distinct melanoma models (YUMMER and YUMM) and (2) treatment conditions (untreated, PD1, and CD8). Moreover, to visualize clusters in 2D space, SVM classification was performed. Based on the p-value from ANOVA and the classification accuracy from SVM, the best performing parameter combination for melanoma discrimination included solely the two H-scan parameters. Therefore, from the normalized H-scan % blue and SNR, PCA calculated the first and second principal components, PC1 and PC2, respectively. PC1 is a combined parameter resulted from combining information through multiparametric analysis, and further to visualize the results, SVM classification was applied to PC1 and PC2.

Results

YUMMER vs YUMM

We subcutaneously implanted the murine melanoma cell lines, YUMM or YUMMER, into the mice flanks to generate an animal model of melanoma. A previous study [25] demonstrated that these two cell lines established immunologically distinct TMEs in vivo with YUMMER tumors having significantly more immune cells than YUMM (Fig 5). In Fig 5 (a-d), the smaller and darker cells likely represent immune cells, whereas the larger cells represent tumor and stromal cells. Immunohistological staining for CD45+, a marker for immune cells, confirmed increased CD45+ immune infiltration in YUMMER tumors compared to YUMM tumors Fig 5 (e-f).

Figure 5.

Representative (a-d) histology images stained with hematoxylin and eosin (H&E) and (e-f) CD45 immunohistochemistry (IHC) for (a,c,e) YUMMER and (b, d, f) YUMM tumors. The red ROI boxes in (a) and (b) were magnified into (c) and (d), respectively.

We measured the following five parameters from US data: H-scan % blue, H-scan SNR, B-scan intensity, B-scan SNR, and the attenuation coefficient. To exclude outliers, we only included measurements that are in the range within two standard deviations for each parameter, when analyzing the measures for statistical analysis and classifications. The results are provided in Fig 6, and only the H-scan parameters were able to distinguish the immunologically distinct TMEs, showing significantly different measurements between YUMMER and YUMM tumors: p < 0.01 for H-scan % blue and p < 0.05 for H-scan SNR. According to the Fig 6 (a) H-scan results, YUMMER tumors had higher percentage of blue pixels, meaning that a greater number of smaller scatterers were detected in YUMMER compared to YUMM tumors. This finding is consistent with the histology which showed higher infiltration of smaller immune cells in YUMMER tumors (Fig 5).

Figure 6.

Estimated ultrasound parameters to differentiate YUMMER and YUMM tumors. (a) H-scan % blue. (b) H-scan SNR. (c) B-scan intensity [dB]. (d) B-scan SNR. (e) Attenuation coefficient [dB/MHz/cm]. * p-value with the following notations: * p < 0.05, ** p < 0.01, *** p < 0.001, **** p < 0.0001, n.s. not significant.

Fig 7 shows representative B-scan and H-scan images for YUMMER and YUMM tumors. The YUMMER tumor H-scan image clearly showed more blue pixels than the YUMM tumor. Due to the increase in smaller immune cells in YUMMER tumors, their H-scan results showed more blue than the YUMM lesions. Also, blue pixels tended to be found near the boundary, which may represent peritumoral infiltration of immune cells, a common clinical phenomenon present in solid tumors.

Figure 7.

Representative B-scan and H-scan images for YUMMER and YUMM tumors. Blue and red colors represent smaller and larger ultrasound scatterers, respectively.

Treatment response

Fig 8 (a) shows the 5 parameter measurements used to detect treatment response. We investigated three groups: untreated (UT), anti-PD-1 antibody treated, and CD-8 depleted groups. Earlier studies demonstrated that anti-PD-1 treatment increased percent of CD45+ immune infiltration from approximately 20% - 30% to 30% - 50% whereas CD-8 depletion decreased immune infiltration in YUMMER tumors [25–28]. In contrast, immune infiltration into YUMM tumors was not altered by either treatment.

Figure 8.

Ultrasound parameters to investigate treatment response in YUMMER (row a) and YUMM (row b) tumors. The parameters were H-scan % blue, H-scan SNR, B-scan intensity [dB], B-scan SNR, and attenuation coefficient [dB/MHz/cm]. The H-scan parameters were combined using principal component analysis (row c). The combined parameter (PC1) differentiated YUMMER and YUMM tumors, detected treatment response in YUMMER tumors, and showed no response for YUMM tumors. * UT: untreated group, PD1: anti-PD-1 treatment, CD8: CD8 depletion. p-value with the following notations: * p < 0.05, ** p < 0.01, *** p < 0.001, **** p < 0.0001, n.s. not significant.

As shown in the leftmost plot in Fig 8 (a), H-scan % blue can detect anti-PD-1 treatment response with a statistically significant difference (p < 0.01). Anti-PD-1 treatment caused a red shift (lower % blue), indicating that there was an increase in larger US scatterers due to the increase in smaller immune cells. While this may appear inconsistent with histology, clusters of smaller immune cells can convalesce and be read as larger US scatterers in an US echo signal. According to Fig 8 (a) and (b) measurements, these individual parameters tend not to be able to show differences between the three groups. Thus, multiparametric analysis was performed and the feature selection process found that combining only the two H-scan parameters most accurately described the treatment response. Fig 8 (c) provides the combined H-scan parameter of PC1 derived from the two H-scan parameters. PC1 shows a significant difference between YUMMER and YUMM tumors (p < 0.05). Moreover, PC1 can detect both anti-PD-1 and CD-8 depletion treatment responses in YUMMER tumors, as demonstrated in the earlier study [25]. According to the study, YUMM was non-responder to PD1 and CD8 treatment, which was also found using PC1.

Fig 9 displays B-scan and H-scan images for the responder YUMMER tumors. H-scan can show slightly different colors for the cases, illustrating the immunologically distinct TMEs in YUMMER tumors secondary to immune modification treatments.

Figure 9.

Representative B-scan and H-scan images of YUMMER untreated, PD1 treated, and CD8 treated cases.

SVM classification

The H-scan analysis is capable of differentiating YUMMER and YUMM tumors, and it can also detect treatment response for the responder YUMMER tumors. Fig 10 shows clustering and classification results in multiparametric space. SVM accurately classified YUMMER and YUMM tumors (Fig 10 (a)) and also distinguished treatment groups (Fig 10 (b)). YUMMER and YUMM tumors can be classified by SVM with 83.3% accuracy. Untreated, anti-PD-1 treated, and CD8-depleted YUMMER tumor groups showed 88.2% classification accuracy. However, YUMM tumors do not respond to the treatments; measurement in 2D space showed overlaps between groups. SVM was not applied to YUMM tumor measurements since the classifier cannot define meaningful hyperplanes without clustered classes.

Figure 10.

SVM classification. (a) Classification between different YUMMER and YUMM form distinct TMEs with different intratumoral immune cell infiltration. (b) Classification between untreated and immunotherapies. * UT: untreated group, PD1: PD1 anti-PD-1 treatment, CD8: CD8 depletion.

Discussion

Previous QUS and H-scan studies aimed to discriminate distinct diseases or differentiate diseased from normal condition [2, 5, 6, 13, 14, 16, 17, 19–21, 29, 30]. However, this study demonstrated that the H-scan is also sufficiently accurate to detect subtle changes within the same type of diseased tissues; the H-scan detected TME heterogeneity between different melanoma tumors. When comparing different types of diseases, morphological pattern differences were likely found in histology images. For instance, fibrosis developed fibrotic septae [31], whereas steatosis generated circle-shaped fat accumulation [14]. In contrast, melanoma metastasis, recapitulated by YUMM and YUMMER murine tumors, may be grossly morphologically identical except for slightly different cellular distributions or sizes measured on a micron scale (Fig 5). These differences are thus more challenging to detect compared to distinct morphologies of different diseases. In fact, these TME changes were unable to be detected by the widely used B-scan intensity parameters and attenuation coefficient measurements (Fig 6 and 8). However, H-scan was capable of detecting these cellular differences. Therefore, this study demonstrated the potential of H-scan to be used as a more precise diagnostic tool in assessing subtle changes in the TME, since we revealed the statistically significant differences with p-values < 0.05 and high classification accuracies > 80%. However, as seen in Fig. 10, clusters of YUMMER and YUMM are likely to be overlapped, and YUMMER cases were located near the boundary of SVM hyperplane due to its small number of samples < 10. To define more accurate hyperplane, future study requires to include larger dataset.

When measuring US features, we have averaged measurements within the entire melanoma lesion, but future studies might include evaluation with localized resolution. Immune cell infiltration can preferentially localize to a specific area or subsequently migrate, causing different immune cell distribution within a lesion. Furthermore, genetic differences between melanoma metastases within an individual host can result in distinct TMEs. These intratumoral and intertumoral heterogeneity (1) cause difficulties in accurate diagnosis following a single biopsy, (2) obscure precise immunotherapy selection, and (3) result in differential treatment responses [24, 25, 32]. Therefore, non-invasive and localized measurement and imaging can contribute to melanoma treatment selection and monitoring. The non-invasive H-scan analysis has pixel-wise resolution. Thus, it is necessary to evaluate H-scan accuracy by comparing the same sections imaged by H-scan with histology to verify that the H-scan can precisely illustrate the heterogeneously distributed immune cells within a lesion and between two different lesions.

The number of immune cells for YUMMER tumors were reported in the following order: CD8-depleted < control < anti-PD-1 treated lesions. In contrast, H-scan % blue showed the following order: anti-PD-1 treated < CD8 depleted < control YUMMER tumors. H-scan measurement corresponds to US scatterer sizes; smaller US scatterers result in higher H-scan % blue. The size of most cancer cells has been reported as 10–20 $\mu m$, and lymphocytes are generally 5–10 $\mu m$ in diameter size [33–38]. Thus, we can expect the smaller immune cells would result in higher % blue. This was seen in in Fig 7; YUMMER tumors showed higher immune cell infiltration than YUMM tumors, resulting in more blue pixels in the YUMMER H-scan image. In addition, when comparing the measurements between the CD8-depleted, which eliminated CD8+ lymphocytes, and the control YUMMER tumors, control YUMMER tumors resulted in higher H-scan % blue, as expected. However, although anti-PD-1 treated YUMMER tumors had greater immune cell infiltration, they showed the lowest blue percentages, meaning US scatterers were revealed to be larger with a higher % red. Due to clustering of immune cells, US scatterers in anti-PD-1 treated YUMMER scans can appear larger than single immune and tumor cells. Moreover, the US wavelength was 38.5 $\mu m$, and the transducer manufacturer (FUJIFILM VisualSonics, Inc., Toronto, Canada) has reported its axial resolution is 40 $\mu m$. The immune cell size (approximately 5–10 $\mu m$) in Fig 5 was found to be smaller than the resolution. Thus, US cannot differentiate single immune cells but only detect changes caused by immune infiltration present in tumor lesions. We can infer that the relatively sparse distribution of the small immune cells can produce more frequent intensity changes for the reflected US echo, acting as smaller US scatterers, whereas the relatively dense distribution of the immune cells with clusters may act as larger US scatterers. Future work could investigate the US scatterer size changes detected by H-scan with varying small cell density. H-scan trajectories along with gradual increase in immune cell infiltration can be studied to monitor melanoma progression.

In this study, five US parameters were extracted from H-scan imaging, conventional B-scan imaging, and the attenuation estimation. The H-scan measures detected TME heterogeneity more sensitively than the other parameters. Consistently, previous studies comparing H-scan measures with other parameters, such as attenuation, echo intensity, and shear wave elastography parameters, have reported that the H-scan analysis resulted in the most accurate tissue classification [19, 20, 39]. However, other parameters were still able to differentiate significant differences between diseases (e.g., normal vs. stage 4 steatosis), for example, B-scan, shear wave elastography (SWE), and H-scan can be used to diagnose severe steatosis [39]. In this study, the only parameters sensitive enough to detect TME heterogeneity were the H-scan parameters (the heterogeneity was not detectable by attenuation or backscattering intensity) because the H-scan utilizes frequency information which reflects subtle changes of scatterers. Further, other QUS parameters can be explored to assess melanoma; based on this study, spectral-based approaches, such as spectral slope (SS) and mid-band fit (MBF), would contribute more than intensity-based approaches, such as Nakagami parameter. However, the classification performance of parameters can vary depending on the tissue characteristics we aim to detect. When classifying breast cancer, lesion boundary shapes are one of the significant factors, which cannot be measured by the H-scan detecting scattering signatures. A breast cancer study reported that boundary shape parameters played the most crucial role in identifying small-sized breast lesions [21]. Therefore, we recommend selecting the H-scan analysis as the most precise metric when detecting US scattering changes. Future study can also include all parameters from H-scan, QUS, and SWE, and compare the performance.

Conclusion

The H-scan analysis was capable of discriminating tumor microenvironmental heterogeneity using the murine melanoma models of YUMMER and YUMM tumors. Moreover, the H-scan was sufficiently precise in monitoring changes in immune infiltration following immunomodulatory therapies. These TMEs were not distinguished by other widely used US features. Overall, the H-scan approach is promising in the identification of subtle pathological changes and has the potential to detect intertumoral heterogeneity in melanoma metastases within an individual host, which could ultimately help the individualize immunotherapy selection in metastatic melanoma.

Data Availability

Archived data is available upon reasonable request.