Abstract

   The present study investigated ultrasound (US) phenotypes reflecting
   prostate cancer (PCa)‐related genetic mutations. Herein, integration of
   radiotranscriptomic data, US and contrast‐enhanced ultrasound (CEUS)
   radiomic images, and RNA sequencing was performed with the aim of
   significantly improving the accuracy of PCa prognosis. We performed
   radiotranscriptomic analysis of clinical, imaging, and two genomic
   (mRNA and microRNA expression) datasets from 48 and 22 men with PCa and
   benign prostatic hyperplasia (BPH), respectively. Twenty‐three US
   texture features and four microvascular perfusion features were
   associated with various patterns of 52 differentially expressed genes
   related to PCa (p < 0.05); 17 overexpressed genes were associated with
   two key texture features. Twelve overexpressed genes were identified
   using microvascular perfusion features. Furthermore, mRNA and miRNA
   biomarkers could be used to distinguish between PCa and BPH. Compared
   with RNA sequencing, B‐mode and CEUS features reflected genomic
   alterations associated with hormone receptor status, angiogenesis, and
   prognosis in patients with PCa. These findings indicate the potential
   of US to assess biomarker levels in patients with PCa.

   Keywords: prostate cancer, radiomics, radiotranscriptomics, ultrasound

Short abstract

   1. Our radiotranscriptomics investigations of PCa provided three
   ultrasound key features identified with radiomics: two texture features
   and one microvascular perfusion feature. 2. The following co‐regulated,
   differentially expressed mRNAs and microRNAs were found in three
   datasets related to prostate cancer: SNAI2, CAV1, RD5A2, RND3, and
   miR205‐5p, miR205‐3p, miR221‐3p. 3. The enriched functions in the three
   key features of the significant genes included negative regulation of
   epithelial cell proliferation, the androgen metabolic process, male
   gonad development, and the lipid biosynthetic process.

1. INTRODUCTION

   The worldwide prostate cancer (PCa) burden is expected to grow to
   approximately 2.3 million new cases and 740,000 deaths by 2040 due to
   the aging of the population.[44] ^1 The incidence of PCa is rapidly
   increasing in response to economic development, prolonged life
   expectancy, and adaptation to Western lifestyles in Asia.[45] ^2 The
   complexities of primary PCa diagnosis are rooted in the multifocal
   nature of the disease. Numerous reports have documented that >80% of
   primary PCa cases show multiple topographically and
   histomorphologically distinct tumor foci.[46] ^3 Next‐generation
   sequencing has shown high levels of genomic diversity among patients
   (interpatient heterogeneity), within a given primary tumor
   (intratumoral heterogeneity), and between distinct tumor lesions and
   different metastatic sites (intertumoral heterogeneity).[47] ^4

   Radiomics can be used to mine high‐throughput quantitative and
   noninvasive imaging features, such as heterogeneity or asymmetrical
   enhancement, to improve cancer diagnosis and treatment.[48] ^5
   Radiotranscriptomics[49] ^6 is a branch of radiogenomics that combines
   “radiomics” and “transcriptomics”, including the expression level of
   messenger RNA (mRNA), microRNA (miRNA), and long noncoding RNA, which
   can also be useful in the search for potential tumor biomarkers.[50] ^7
   The rationale behind the implementation of a new multiparametric
   approach (i.e., combining the information from complementary biomarkers
   such as tissue texture, elasticity, or perfusion, with both magnetic
   resonance imaging and ultrasound) is to overcome the complexity of PCa
   as a multifocal and heterogeneous disease. Several advances in US
   techniques have improved the detection of PCa. Identifying imaging
   features associated with a high risk of clinically significant disease
   that are functionally related to defined molecular alterations will
   ultimately allow the detection of molecular heterogeneity within tumors
   and provide an exciting new intersection between imaging and
   transcriptomics. Combined with molecular imaging, these methods will
   significantly improve tissue sampling and increase the accuracy of
   precision medicine for treating localized PCa.

   We assumed that the morphology and vascular phenotypes visualized with
   US could be correlated with the expression of specific genes reflecting
   the growth or angiogenesis in PCa and that radiotranscriptomic
   signatures at the imaging level could capture the underlying intratumor
   heterogeneity at the molecular level. Several radiotranscriptomic[51]
   ^8 , [52]^9 , [53]^10 investigations have described the association
   between PCa features and genetic alterations using magnetic resonance
   imaging but rarely US.[54] ^11 This study underscores the potential of
   combining traditional clinical features with radiotranscriptomics. By
   identifying specific biomarkers, our study suggests that the accuracy
   of predicting prostate cancer (PCa) can be enhanced. Our findings
   focused on the radiotranscriptomics‐based identification of new
   biomarkers that can distinguish prostate cancer from benign prostatic
   hyperplasia.

2. MATERIALS AND METHODS

2.1. Workflow of transcriptomics and radiomic data approach

   The radiotranscriptomic workflow is shown in Figure [55]1. It
   illustrates our integrated approach to jointly analyzing clinical,
   imaging, and transcriptomic data from patients with PCa. Radiomic
   analysis included texture and microvascular perfusion features, which
   were used to extract imaging features from the PCa tumor region.
   Following the guidelines and reporting recommendations of IBSI helps
   ensure that the texture features are reproducible and comparable.[56]
   ^12 , [57]^13 The types, parameter settings, and value range of gray
   intensity and texture features in this study were implemented using a
   Medical Interaction Toolkit (MITK) ([58]https://docs.mitk.org/2021.02/)
   for the manual delineation of tumor boundaries and identification of
   tumor regions of interest (ROI) for radiomics analysis.[59] ^14 This
   ensured the reproducibility and comparability of the results. The
   specific methods are presented in Supplemental Materials. The
   microvascular perfusion features of CEUS images and video clips were
   analyzed using the QontraX software 4.00 (Esaote), and whole‐image
   parameter analysis was performed on a manually drawn reference plane
   (pink panel) for perfusion within a specific ROI in the entire region
   of the selected in‐frame tumor. Tissues from the identified targets
   were obtained for the analysis of differentially expressed genes (DEGs)
   using next‐generation sequencing (green panel) (Gene Denovo
   Biotechnology Co., Ltd). The final step was linking the radiomic
   information to the transcriptome data to identify high‐performance
   candidate imaging features (yellow panel).

FIGURE 1.

   FIGURE 1
   [60]Open in a new tab

   Experimental design. (A) Clinical data (blue panel); (B) Imaging data
   (radiomic analysis) (pink panel); (C) Differentially expressed mRNA and
   miRNAs (green panel); (D) The final step is to link radiomic
   information to transcriptome data (yellow panel).

2.2. Study participant characteristics

   From September 2020 to July 2021, 70 consecutive participants with
   suspected PCa were scheduled for US‐guided biopsy; 22 tissue samples
   were benign and 48 were malignant. These were used to evaluate the
   B‐mode and CEUS values to distinguish benign from malignant masses.
   B‐mode and CEUS values were used to evaluate the benign and malignant
   masses. The inclusion criteria included: (i) patients with clinical
   symptoms (frequent urination, urgency, or dysuria pain) or
   prostate‐specific antigen (PSA) level >4 ng/mL; (ii) transrectal B‐mode
   US and CEUS examinations were performed before US‐guided biopsy; (iii)
   pathological results were confirmed by biopsy. The exclusion criteria
   included: (i) allergy to US contrast agents (n = 1); (ii) inadequate
   image quality of US and CEUS scans (n = 21); or (iii) inadequate gene
   sequencing samples (n = 3).

2.3. US imaging acquisition

   Voxel size is one of the most important imaging parameters that varies
   significantly depending on the modality, vendor, and imaging protocols
   used. To reduce the influence of US resolution on voxels, we uniformly
   fixed the imaging parameters of ultrasound images. We used a MyLab
   Twice US system (Esaote SpA) with an EC1123 transrectal probe
   (frequency: 3–9 MHz; contrast frequency: 3–6 MHz), mechanical index of
   0.04–0.13, gain adjustment: 66%, depth: 53 mm, and using X‐View
   adaptive imaging technology (automatic removal of intelligent speckle
   noise in the whole process, including intelligent sound beam
   adjustment, signal speckle noise suppression, pixel optimization
   adjustment, and other technologies for improved imaging quality).
   Abnormal echo patterns (calcifications, cysts, and hypoechoic lesions)
   of the prostate were recorded. B‐mode and CEUS examinations were
   performed by two radiologists who analyzed the imaging phenotypes. A
   senior sonographer (>10 years of experience) performed manual
   delineation of the lesion boundary and determined that the ROI
   minimized the influence of subjective differences. Another senior
   sonographer (>10 years of experience) confirmed the delineation of the
   lesion area. In cases of disagreement, the two sonographers discussed
   the issue and reached an agreement. If a patient had multiple
   suspicious lesions, the largest lesion was used as an ROI for further
   evaluation. The US contrast agent SonoVue (Bracco SpaA) consisted of
   microbubbles filled with 2.5 μm sulfur hexafluoride. For CEUS
   examination, a rapid intravenous bolus (4.8 mL) of SonoVue was
   administered, followed by irrigation with 5 mL of normal saline.
   Continuous real‐time digital imaging was performed immediately after
   injection for 5 min. The imaging data were retrospectively analyzed.

2.4. Prostate biopsy procedure and pathology examination

   The US instrument scanner and transrectal probe used in prostate biopsy
   were consistent with TRUS. Biopsy was performed using an 18‐G biopsy
   gun (Bard Biopsy Systems, Tempe, AZ, USA) with a length of 18 mm and a
   penetration depth of 22 mm. In this study, we used multiple MRI scans
   to determine whether the tumor protruded from the exocapsular growth or
   invaded the seminal vesicle prior to the biopsy procedure. Transrectal
   ultrasound (TRUS) examinations were performed by employing blanket
   transperineal biopsies to reduce infection‐related complications. The
   lesion selected as the ROI was then biopsied. All patients underwent
   the “12 + X” biopsy, a targeted biopsy for suspicious areas (combined
   with B‐mode and CEUS) based on a 12‐core transrectal systematic biopsy
   performed by one radiologist with more than 10 years of biopsy
   experience. The needles were injected into the cognitive suspicious
   targeted nodules. The cognitive targeted biopsy was performed as
   follows: first, the radiologist reviewed the MRI results; and second,
   the radiologist used the US information to perform the targeted biopsy
   for the most remarkable suspicious nodules guided by US images. In our
   study, tissue biopsy specimens were collected with two biopsy cores in
   each region. One specimen was used for pathological diagnosis, while
   the other was flash frozen in liquid nitrogen for next‐generation
   sequencing (mRNA and miRNAs). We recorded spatial locations of each
   biopsy site in detail and used the size and pathological findings to
   determine the location and nature of the lesion, thus ensuring that the
   description of pathological locations matched the relevant lesions on
   TRUS images. Regarding pathological results, all histopathologic
   specimens were labeled as cancer or benign prostatic hyperplasia
   lesions using a combination of primary and secondary Gleason score by a
   pathologist. For each biopsy core, the maximum Gleason score was
   recorded.

2.5. mRNA and miRNA sequencing and analysis

   Total RNA was prepared using the Trizol reagent (Invitrogen). The
   quality of RNA samples was assessed with an Agilent 2100 Bioanalyzer
   (Agilent Technologies). Total RNA (1 μg per sample) was used for the
   construction of a sequencing library. Eukaryotic mRNAs were enriched by
   Oligo(dT) beads and by removing prokaryotic mRNAs and rRNAs using the
   Ribo‐ZeroTM Magnetic Kit (Epicenter). Enriched mRNAs were then
   fragmented using a fragmentation buffer and reverse transcribed into
   cDNAs with random primers. After synthesis, cDNAs were purified using
   the QiaQuick PCR extraction kit (Qiagen), ligated to Illumina
   sequencing adapters, selected by gel electrophoresis, and amplified by
   PCR. Subsequently, cDNA libraries were sequenced using the Illumina
   Novaseq 6000 by Guangzhou GENE DENOVO Biology (Co. Ltd.). For miRNA
   sequencing, RNA fragments of 18–30 nucleotides in length were enriched
   by polyacrylamide gel electrophoresis (PAGE). After adding 3′ and 5′
   adapters, samples were subjected to RT‐PCR, and PCR products with
   140–160 bp size were enriched to generate a cDNA library and sequenced
   using the Illumina Novaseq 6000.

   DEGs were assessed by analyzing differential RNA expression between the
   two groups. Transcripts with a p‐value <0.05 and an absolute fold
   change of ≥2 were considered differentially expressed. Namely, miRNAs
   that exhibited an absolute fold change of ≥1.5 and a p‐value <0.05 were
   classified as differentially expressed. Correlations between US imaging
   phenotypes and gene expression were analyzed using Pearson's
   correlation coefficients and the R package psych (Version 2.2.3)
   ([61]http://personality‐project.org/r/psych/HowTo/getting_started.pdf).
   The correlations between the miRNAs and their co‐expressed mRNA were
   analyzed using a Sankey diagram. Cytoscape software (version 3.7.1;
   [62]http://www. [63]cytoscape.org/) and the ggalluvial R package
   ([64]https://cran.r‐project.org/web/packages/ggalluvial/vignettes/ggall
   uvial.html) were used to visualize the Sankey diagram. Gene Ontology
   (GO) ([65]http://geneontology.org/) and Kyoto Encyclopedia of
   Enrichment Genes and Genomes (KEGG) ([66]https://www.genome.jp/kegg/)
   pathway analyses were performed using the GO and KEGG databases. The
   calculated p‐value was corrected for false discovery rate, using a
   false discovery rate of 0.05 as a threshold. The real‐time interactive
   online data analysis platform OmicSmart ([67]http://www.omicsmart.com)
   was used for bioinformatics analysis. The RNA‐seq data generated in
   this research have been placed in the NCBI Sequence Read Archive
   database under the accession code PRJNA916752
   ([68]http://www.ncbi.nlm.nih.gov/bioproject/PRJNA916752).

2.6. Statistical analysis

   In the model validation phase, we tested and compared the performance
   of three machine learning predictive models (random forest, naïve
   Bayes, and SVM) with four datasets, including clinical features,
   molecular biomarkers, radiomics features, and the combined datasets.
   Data were analyzed using GraphPad Prism, v9.0.2 (GraphPad Software),
   and summarized as the mean ± SD. Correlation analysis was performed
   using Pearson analysis. Unpaired two‐tailed Student's t‐tests or
   analysis of variance were used to compare two or more groups, as
   indicated in the figure legends.

3. RESULTS

3.1. Clinical and transcriptomic data collected from patients with Pca and
BPH enabled radiotranscriptomic analysis

   We processed the clinical data (age, total PSA, ratio of free to total
   PSA, prostate volume, PSA density, prostate biopsy pathology, biopsy
   cores, positive biopsy cores, Gleason score, and pathological stage)
   and two types of transcriptomic (mRNA and miRNA) data collected from 48
   patients with primary PCa tumors and 22 patients with benign prostatic
   hyperplasia (BPH). PSA density was calculated as total PSA/prostatic
   volume. Table [69]1 summarizes the clinical and pathological
   characteristics of participants.

TABLE 1.

   Characteristics of the patient group.
   Parameters Malignant Benign p‐value
   Number of patients, n 48 22
   Age, year 69.67 ± 7.64 (54–88) 66.36 ± 7.35 (52–78) 0.09
   Total PSA, ng/mL 15.67 ± 9.56 (4.99–49.24) 10.88 ± 5.21 (4.33–20.05)
   <0.05
   fPSA/tPSA 0.11 ± 0.08 (0.01–0.33) 0.16 ± 0.05 (0.06–0.28) <0.01
   TRUS volume, mL 43.71 ± 20.09 (13–105) 66.14 ± 31.91 (25–147) <0.01
   PSA density, ng/mL^2 0.49 ± 0.56 (0.04–1.17) 0.19 ± 0.14 (0.04–0.45)
   0.02
   Biopsy cores 11.73 ± 1.32 (8–13) 12 0.34
   Positive biopsy cores 7.25 ± 3.69 (1–12) 0 <0.01
   Lesion Gleason scores
   3 + 3 = 6 3 0
   3 + 4 = 7 6 0
   4 + 3 = 7 16 0
   4 + 4 13 0
   4 + 5 9 0
   5 + 5 1 0
   Pathological stage
   T1 0 0
   T2 34 0
   T3 13 0
   T4 1 0
   N0 0 0
   N1 3 0
   N2 0 0
   N3 0 0
   [70]Open in a new tab

   Abbreviations: fPSA, free PSA; PSA, prostate‐specific antigen; tPSA,
   total PSA; TRUS, transrectal ultrasound.

3.2. B‐model, CEUS imaging, and machine‐learning methods revealed 23
significant texture features

   The importance score for the texture features was calculated, as shown
   in Figure [71]S1A. In both the B‐mode and CEUS images, 368 radiomic
   texture features were filtered down to 23 significant texture features.
   These were used to build a classification model based on multiple
   logistic regression, a random forest model, and a support vector
   machine. For B‐mode images, when the number of features used to
   construct the classification model is greater than five, the accuracy
   of the classification model can exceed 0.95 or even be close to 1. For
   CEUS images, the accuracy of the classification model was close to 1
   (Figure [72]S1B–D).

3.3. Correlation analysis of texture and microvascular features identified 6
upregulated genes and 9 downregulated genes as markers of Pca progression

   A total of 4080 DEGs were identified, including 1231 upregulated genes
   and 2849 downregulated genes, in PCa and BPH. In order to narrow the
   scope of the study, DO database was used for enrichment [DO (Disease
   Ontology), a database describing gene function and disease related],
   and 52 significantly DEGs were found to be related to PCa A total of 52
   DEGs associated with PCa were identified. A full list of the 52 DEGs is
   presented in Table [73]S2. To determine whether DEGs might be coupled
   based on radiomic texture features, we used pie‐chart mapping to
   understand all pairwise correlations between significant texture
   features, microvascular perfusion, and genes (Figure [74]2). Five
   significant texture features were highly correlated with different
   genes. DEGs associated with significant texture features and
   microvascular perfusion features are seen in Tables [75]S3 and [76]S4,
   respectively. Gene expression data according to the cd1‐gray‐level
   route length matrix (GLRLM)‐gray‐level variance (GLV) in B‐mode US
   using the heat map (Figure [77]3A) and volcano plots (Figure [78]3B)
   indicated that the cd1‐GLRLM‐GLV of PCa was associated with the
   upregulation of fatty acid synthase (FASN) and Fragile Histidine Triad
   (FHIT). This association is relevant to PCa in terms of lipid
   metabolism, androgen receptor status, drug resistance, proliferation,
   apoptosis, and the downregulation of steroid 5 alpha‐reductase 2
   (SRD5A2), Snail family transcriptional repressor 2 (SNAI2), Rho family
   GTPase 3 (RND3), and caveolin‐1 (CAV1).[79] ^15 , [80]^16 The latter
   are associated with PCa based on genes reported to be relevant in
   androgen receptor signaling inhibition, castration‐resistant PCa, and
   the proliferation and migration of PCa.[81] ^17 Gene expression data
   according to cd1‐gray‐level size zone matrix (GLSZM)‐small zone low
   gray‐level emphasis (SZLGE) at CEUS using the heat map (Figure [82]3C)
   and volcano plots (Figure [83]3D) indicated that the cd1‐GLSZM‐large
   zone low gray‐level emphasis (LZLGE) is associated with the
   downregulation of WNT inhibitory factor 1 (W1F1), SNAI2, SRD5A2,
   Pleiotrophin (PTN), RND3, Laminin subunit beta 1 (LAMB1), Sushi repeat
   containing protein X‐linked (SRPX), CAV1, Matrix metallopeptidase 16
   (MMP16), and Cyclin D2 (CCND2) genes, which are associated with the
   promotion of cell proliferation and metastasis based on the genes
   reported to be relevant to PCa.[84] ^18

FIGURE 2.

   FIGURE 2
   [85]Open in a new tab

   Summary of radiotranscriptomic correlations. (A) Colored circles refer
   to 23 significant texture features of B‐mode ultrasound (US) and genes;
   (B) Colored circles refer to 23 significant texture features of
   contrast‐enhanced ultrasound (CEUS) and genes; (C) Colored circles
   refer to 4 significant microvascular perfusion features of CEUS images
   and genes. The left half of the pie charts represents the features of
   radiomics, and the right counterpart represents the related genes.
   Ribbons connecting texture features and genes refer to significant
   correlations between them. Ribbon colors refer to the correlation
   coefficient sign (blue, negative; red, positive).

FIGURE 3.

   FIGURE 3
   [86]Open in a new tab

   Radiotranscriptomic correlations according to key texture features with
   B‐mode ultrasound (US) and contrast‐enhanced ultrasound (CEUS). (A) A
   heat map image demonstrates six significant differentially expressed
   genes (DEGs) according to the cd1‐gray‐level route length matrix
   (GLRLM)‐gray‐level variance (GLV). The columns represent 70 individuals
   with either prostate cancer (PCa) (orange) or benign prostatic
   hyperplasia (BPH) (blue) and the rows represent six individuals with
   significant DEGs. The color key indicates the degree of DEGs in either
   direction: upregulation (orange) or downregulation (blue); (B) A
   volcano plot shows significant DEGs with cd1‐GLRLM‐GLV (p < 0.05).
   Purple circles represent significant DEGs between PCa and BPH with
   p < 0.05 and log[2]FC >2.0 or < −2.0; (C) A heat map shows 10
   significant DEGs according to the cd1‐gray‐level size zone matrix
   (GLSZM)‐large zone low gray‐level emphasis (LZLGE). The columns
   represent 70 individuals with either PCa (orange) or BPH (blue), and
   the rows represent 10 individuals with significant DEGs. The color key
   indicates the degree of DEGs in either direction: upregulation (orange)
   or downregulation (blue); (D) A volcano plot demonstrates the
   significant DEGs with cd1‐GLSZM‐LZLGE (p < 0.05). Purple circles
   represent significant DEGs between PCa and BPH with a p‐value <0.05 and
   a log[2]FC >2.0 or < −2.0.

   Analysis of CEUS images revealed that PCa did not penetrate
   (Figure [87]4A) the vessels in an 85‐year‐old man, whereas BPH was
   found to partially penetrate (Figure [88]4B) the vessels in a
   79‐year‐old man. Three‐dimensional images revealed an enhancement in
   microvascular perfusion and regional blood flow (RBF) in the ROI of PCa
   and BPH (Figure [89]4C,D). Gene expression data based on microvascular
   perfusion features using heat maps (Figure [90]4E) and volcano plots
   (Figure [91]4F) indicated that the RBF was associated with the
   upregulation of Solute carrier organic anion transporter family member
   1B3 (SLCO1B3), Alpha‐methylacyl‐CoA racemase (AMACR), SIM2,
   Glucosaminyl (N‐acetyl) transferase 1 (GCNT1), PTN, and Catenin delta 2
   (CTNND2) genes, which are relevant to androgen deprivation therapy
   resistance, early diagnosis biomarkers, tumor cell transcription
   factors, tumor aggressiveness, prometastasis, cell proliferation, and
   the migration and differentiation of PCa.[92] ^19 , [93]^20 , [94]^21 ,
   [95]^22 , [96]^23 They are also relevant to the downregulation of
   SNAI2, Mal, T‐cell differentiation protein (MAL), RND3, LAMB1, CAV1,
   MMP16, SRPX, ITGB3, and TGF‐beta signaling pathway (BMP2) genes, which
   are reported to be relevant to metastasis, cancer progression,
   motility, invasion into the surrounding extracellular matrix, and the
   proliferation, and migration of PCa.[97] ^24 , [98]^25

FIGURE 4.

   FIGURE 4
   [99]Open in a new tab

   Radiotranscriptomic correlation according to the presence of
   microvascular perfusion in contrast‐enhanced ultrasound (CEUS) in 70
   participants. (A) CEUS image (orange border) shows prostate cancer
   (PCa) in an 85‐year‐old man that does not penetrate the vessels; (B)
   CEUS image (blue border) shows benign prostatic hyperplasia (BPH) in a
   79‐year‐old man with vessels that are partially penetrated (arrows). In
   the CEUS images, yellow circles are the regions of interest for
   measuring quantitative parameters; (C, D) Three‐dimensional images
   reveal enhancement with microvascular perfusion regional blood flow
   (RBF) in the region of interest of PCa and BPH tissues; (E) A heat map
   image demonstrates 16 DEGs according to the RBF. The columns represent
   70 individuals with either PCa (orange) or BPH (blue), and the rows
   represent 16 individuals with DEGs. The color key indicates the degree
   of DEGs in either direction: upregulation (orange) or downregulation
   (blue); (F) A volcano plot shows 16 DEGs with RBF (p < 0.05). Purple
   circles represent DEGs between PCa and BPH with p < 0.05 and
   log[2]FC >2.0 or < −2.0.

3.4. 53 and 63 miRNA gene networks, consisting of 9 and 14 mRNAs, were
associated with the texture feature cd1‐GLSZM‐SZLGE and microvascular feature
of PCa, respectively

   In order to recognize the effect of radiotranscriptomics on
   miRNA‐mediated mRNAs in PCa, we built a regulatory network based on
   previously reported data. The correlation between the radiomic‐related
   features of the miRNA–mRNA regulatory network was analyzed using the
   Sankey diagram. As shown in Figure [100]5A–C, the core miRNA‐gene
   network comprised five mRNAs associated with the texture feature
   cd1‐GLRLM‐GLV of PCa and 33 miRNAs targeting these genes, among which
   two or more miRNAs might target the same mRNA. A total of 53 and 63
   core miRNA gene networks, consisting of 9 and 14 mRNAs, were associated
   with the texture feature cd1‐GLSZM‐SZLGE and the microvascular features
   of PCa, respectively.

FIGURE 5.

   FIGURE 5
   [101]Open in a new tab

   Sankey diagram for the co‐network of transcriptomics‐related mRNA and
   its co‐expressed microRNA (miRNA). (A) Core miRNA–gene network
   consisting of five mRNAs associated with the texture feature
   cd1‐gray‐level route length matrix (GLRLM)‐GLV of PCa and 33 miRNAs
   targeting the genes; (B) Core miRNA–gene network consisting of five
   mRNAs associated with texture feature cd1‐gray‐level size zone matrix
   (GLSZM)‐large zone low gray‐level emphasis (LZLGE) of PCa; (C) Core
   miRNA‐gene network consisting of 16 mRNAs associated with the
   microvascular feature RBF of PCa. The Venn diagram shows the mRNAs and
   miRNAs associated with three key features; (D) Four mRNA co‐expressions
   were obtained, including SNAI2, CAV1, RD5A2, and RND3; (E) Three miRNA
   co‐expressions were obtained, including miR205‐5p, miR205‐3p, and
   miR221‐3p.

3.5. Seven coregulated mRNAs and miRNAs were associated with three key
features, cd1‐GLRLM‐GLV, cd1‐GLSZM‐SZLGE, and RBF, respectively

   The mRNAs and miRNAs associated with the three key features,
   cd1‐GLRLM‐GLV, cd1‐GLSZM‐SZLGE, and RBF, were mapped using a Venn
   diagram to identify the mRNAs and miRNAs shared by all three key
   features, as well as those shared by at least two key features. The
   following co‐regulated DEGs and differentially expressed miRNAs were
   found in the three datasets: SNAI2, CAV1, RD5A2, RND3, miR205‐5p,
   miR205‐3p, and miR221‐3p.

3.6. Pca pathways of differentially expressed genes were implicated in three
key US imaging features

   GO and KEGG pathway analyses were used to explore the enrichment of the
   functions and pathways of key features related to DEGs. The functions
   associated with PCa were annotated primarily with respect to the key
   features of the US images (Table [102]S5). Significantly upregulated or
   downregulated genes with enriched functions were identified in three
   key US imaging features, including negative regulation of epithelial
   cell proliferation, androgen metabolic processes, male gonadal
   development, and lipid biosynthesis processes, which are associated
   with prostate tumorigenesis. Pathway enrichment analysis showed that
   key features related to the PCa pathways of DEGs (such as the Hippo
   signaling pathway and PCa) were implicated in three key US imaging
   features (Table [103]S6).

3.7. New radiotranscriptomic analysis showed significant diagnostic
advantages over traditional radiomic, transcriptomic, and clinical analyses
alone

   We used the receiver operating characteristic (ROC) curves of three
   machine‐learning methods (random forest, naïve Bayes, and SVM) to
   evaluate the dependability of the prediction accuracy of all the
   features in four feature sets (Figure [104]6): (1) the clinical dataset
   (age, PSA, free PSA/total PSA, volume, PSA denticity, and GS); (2) the
   molecular biomarkers (SNAI2, CAV1, RD5A2, RND3, miR205‐5p, miR205‐3p,
   and miR221‐3p); (3) the radiomics features only (texture features and
   microvascular perfusion features); and (4) the combined set. The values
   of the combined set (area under the curve [AUC] = 0.997, 0.987, and
   0.998 for random forest, naïve Bayes, and SVM, respectively) showed
   more accurate (or equivalent) prognostic power than the radiomics
   feature (AUC = 0.998, 0.984, and 0.997), transcriptomics feature (genes
   and miRNAs) (AUC = 0.871, 0.935, and 0.064), or clinical datasets
   (AUC = 0.786, 0.843, and 0.925) (Table [105]S7). These results
   highlighted the possibility of combining the features of different
   datasets to obtain a more accurate diagnosis of PCa.

FIGURE 6.

   FIGURE 6
   [106]Open in a new tab

   Prediction of performance using clinical features, transcriptomics,
   radiomics features, and combined features for predicting PCa. The
   receiver operating characteristic (ROC) curves of three
   machine‐learning methods (random forest, naïve Bayes, and SVM) to
   evaluate the dependability of the prediction accuracy of all the
   features in four feature sets: (A) Random forest mode; (B) Naive bayes
   mode; (C) Support vector machine mode. These curves were selected
   because their area under the curve value was closest to the average
   area value of the respective methods over all 10 runs (see also
   Table [107]S7).

4. DISCUSSION

   TRUS plays an indispensable role in prostate biopsies. However, PCa is
   a multifocal and heterogeneous disease, which has a certain impact on
   TRUS‐guided prostate biopsy[108] ^26 as 30% or more of prostate tumor
   specimens are isoechoic and the probability of cancer being recorded in
   hypoechoic lesions is approximately 50%. Our results included CEUS
   images indicating PCa that did not penetrate the vessels and BPH that
   partially penetrated the vessels. Tumor heterogeneity is difficult to
   assess by irregular sampling or biopsy. The establishment of a
   noninvasive technique to assess the heterogeneity of tumors is a
   clinical necessity. Texture analysis is an image processing method to
   quantify heterogeneity based on pixelwise spatial spread that may be
   missed by the naked eye. Imaging provides quantifiable, clinically
   relevant tumor parameters for the assessment of heterogeneity. The
   application of image signal heterogeneity and pattern recognition
   technology to the analysis of differences that cannot be detected by
   human eyes can greatly enhance and assist diagnosis. TRUS image texture
   features were used to distinguish normal, benign, and malignant
   prostate tissues.[109] ^27 Among the different texture analysis
   methods, the wavelet transform and time‐frequency analysis techniques
   were the most prominent in terms of algorithm efficiency.[110] ^28 We,
   therefore, analyzed 322 wavelet‐filtered features based on US images.
   Maggio et al. proposed a computer‐aided detection of PCa using discrete
   wavelet transform and texture features extracted from TRUS images;
   using texture biomarkers as an intratumoral heterogeneity
   quantification tool held great promise for the early detection of
   PCa.[111] ^29 , [112]^30 We found that cd1‐GLRLM‐GLV was a common key
   texture feature for both B‐mode US and CEUS images and that GLSZM‐SZLGE
   was repeated many times in important CEUS texture features. Therefore,
   our results show that tumor heterogeneity can be quantified by
   radiologic imaging using texture heterogeneity features, which involves
   determining the spatial relationship of image voxel intensities within
   the region of interest. Therefore, texture features may be more
   suitable for detecting heterogeneity of PCa tissue structure in
   imaging.

   Our experimental results demonstrated the importance of optimizing
   texture extraction parameters to improve their predictive value and
   better understand the relationship between texture and biology. Li
   et al. found that a radiomic‐based machine learning model could improve
   the prediction accuracy of clinically significant PCa compared with
   only evaluating the clinical risk factors associated with PCa.[113] ^31
   Sun et al. found that a horizontal comparison of texture features
   (gray‐level co‐occurrence matrix [GLCM] and gray‐level route length
   matrix [GLRLM]) improved the model performance for classifying prostate
   tumor aggressiveness.[114] ^32

   CEUS can reveal abnormal vascular areas of the prostate, thereby
   improving the number of positive biopsies performed.[115] ^33 In a
   recently published study, perfusion‐ and dispersion‐related CEUS
   parameters were successfully combined using a machine learning
   approach.[116] ^34 Our experimental results also confirmed that RBF was
   a key feature with the most DEGs among the four radiomic microvascular
   perfusion features. The latter characteristics can also act as strong
   indicators for PCa and provide insights into its underlying biology.

   We observed strong individual correlations between PCa genes and
   quantitative radiomic features. The Gleason score classifies the degree
   of differentiation and growth pattern of the tumor into five grades;
   the higher the score, the worse the differentiation. Hence, Gleason
   score is closely related to prognosis, and is a tradeoff index for cell
   differentiation and invasion of PCa. FASN and FHIT are commonly
   upregulated during human PCa, and their inhibition was associated with
   promoting cell proliferation and invasion.[117] ^35 , [118]^36 The
   SNAI2 gene, which is silenced in PCa, regulates neuroendocrine
   differentiation, metastasis suppression, and expression of pluripotency
   genes.[119] ^37 , [120]^38 Lower levels of expression of SNAI2 and CAV1
   were associated with higher Gleason scores and advanced pathological T
   and N stages. Thus, SNAI2 and CAV1 can facilitate the identification of
   new molecules or pathways that may be involved in the diagnosis and
   treatment of PCa.[121] ^39 , [122]^40 SRD5A2 is an important enzyme
   involved in androgen metabolism.[123] ^31 RND3, which has been
   identified as a tumor suppressor gene, is constitutively expressed in
   BPH, whereas its expression is downregulated during prostate
   carcinogenesis.[124] ^41 , [125]^42 , [126]^43 In agreement with this
   study, the PCa texture features was associated with the upregulation of
   FASN and FHIT genes, and the downregulation of SNAI2, SRD5A2, RND3,
   CAV1 genes. miRNA is a type of single‐stranded noncoding small RNA
   involved in posttranscriptional regulation of gene expression. Several
   lines of evidence have suggested that miRNAs play key roles in a
   variety of physiological events. Abnormal levels of miRNAs have been
   correlated with the occurrence and development of many diseases,
   including PCa. For instance, the expression of miR‐205 was
   significantly downregulated in advanced PCa.[127] ^44 Studies have
   shown that the miR‐221 gene functions in disease pathogenesis by
   targeting a variety of cellular and molecular pathways.[128] ^45

   Functional analysis of the three key US imaging feature‐specific genes
   revealed diverse functional GO terms and KEGG pathways. Significantly
   upregulated or downregulated genes in three key US imaging features
   were enriched for functions including negative regulation of epithelial
   cell proliferation, androgen metabolic processes, male gonadal
   development, and lipid biosynthesis processes associated with prostate
   tumorigenesis. More gene‐related PCa pathways (such as the Hippo
   signaling pathway and PCa) were involved in the three key US imaging
   characteristics.

   Recent studies have used other machine learning models, such as
   neurofuzzy models (AUC = 0.812)[129] ^46 and artificial networks
   (AUC = 0.695),[130] ^47 to predict PCa staging based solely on clinical
   characteristics. These AUC values were similar to those calculated by
   our machine learning methods using clinical characteristics (0.786,
   0.843, and 0.925, random forest, naïve Bayes, and SVM, respectively).
   Our combined set prediction model showed high potential to yield
   clinically relevant results for characterizing PCa (AUC = 0.997, 0.987,
   and 0.998, for random forest, naïve Bayes, and SVM, respectively).

   Our study had several limitations. First, limiting the occurrence of
   false‐positives in the application of radiomics is important. This
   study uses a variety of methods to minimize the occurrence of false
   positives. We acknowledge that although radiotranscriptomic analysis is
   highly innovative, it is also a truly arduous undertaking as it is
   still in its early stages of development. Future studies are needed to
   validate these findings using a larger cohort of patients. Second, PC
   diagnosis was made based on clinical findings, US imaging findings,
   pathological analysis, and transcriptomic results. We used the ROC
   curves to evaluate the dependability of the prediction accuracy of
   radiotranscriptomics; however, future analysis is needed to determine
   whether radiotranscriptomic analysis can improve the sensitivity and
   specificity of PCa diagnosis with US imaging. Third, radiomics features
   are calculated from matrices that describe various spatial
   relationships between signal intensities of voxels. Voxel size
   resampling is required as a preprocessing step for datasets acquired
   with variable voxel sizes. In future In future 3D (volume)ultrasound
   imaging radiomics studies, we will normalize the size‐dependent
   features by the number of voxels to reduce the dependence on the number
   of voxels in the ROI. Finally, the Gleason score is closely related to
   prognosis and a tradeoff index for cell differentiation and invasion of
   PCa. Our next research goal is to combine stage‐specific genes and
   miRNAs with aggressiveness‐related imaging features, thereby
   constructing radiotranscriptomic signatures for improving the
   prediction accuracy for stages T2 and T3 in patients with PCa.

5. CONCLUSIONS

   In conclusion, when combining traditional clinical features and using
   the power of radiotranscriptomics, the identified biomarkers (SNAI2,
   CAV1, RD5A2, RND3, miR205‐5p, miR205‐3p, and miR221‐3p) could improve
   the accuracy of predicting PCa. This approach can provide unprecedented
   opportunities to improve and support medical decision making, given the
   routine use of imaging in diagnosis and treatment in clinical practice
   worldwide.

AUTHOR CONTRIBUTIONS

   Qian Yang: Conceptualization (lead); data curation (lead); formal
   analysis (lead); funding acquisition (lead); investigation (lead);
   methodology (lead); resources (lead); writing – original draft (lead);
   writing – review and editing (lead). Qiuyang Li: Conceptualization
   (equal); resources (equal). Nan Li: Data curation (lead); software
   (equal). Dingyi Wang: Data curation (equal); formal analysis (equal);
   validation (equal). Shaoxi Niu: Data curation (equal); software
   (equal); validation (equal). Tang Peng: Formal analysis (equal);
   investigation (equal); software (equal). Jing Xiao: Data curation
   (equal); resources (equal). Jiahang Zhao: Data curation (equal);
   resources (equal). Pei Wang: Data curation (equal); resources (equal).
   Yukun Luo: Project administration (lead); supervision (equal). Jie
   Tang: Project administration (lead); supervision (equal).

FUNDING INFORMATION

   This work was supported by grants from the National Natural Science
   Foundation of China (nos. 81471682 and 81801708), China Postdoctoral
   Science Foundation (no. 2021T140795), Natural Science Basic Research
   Program of Shaanxi Province (no.2023‐JC‐QN‐0912), Xi'an Science and
   Technology Plan Project (no. 21YXYJ0134), and Xi'an International
   Medical Center Hospital Level Project (no. 2022QN01).

CONFLICT OF INTEREST STATEMENT

   The authors declare that the research was conducted in the absence of
   any specific financial interests, relationships, or affiliations
   relevant to the subject matter or materials discussed in the manuscript
   (e.g., employment/affiliation; grants or funding; consultancies;
   honoraria; stock ownership or options; expert testimony; royalties; or
   patents filed, received, or pending) that could be construed as
   potential conflicts of interest.

Supporting information

   Data S1.
   [131]Click here for additional data file.^ (194.7KB, docx)

ACKNOWLEDGMENTS