The distinction between ChRCC and RO is crucial due to the significant differences in their prognosis and treatment. RO, being a benign tumour with no risk of metastasis, should be identified in advance to avoid unnecessary treatments. In contrast, ChRCC is a malignant tumour with the potential to spread, requiring more intensive treatment, such as surgical resection and constant monitoring. Misdiagnosing these conditions can lead to overtreatment in the case of RO or inadequate treatment in the case of ChRCC, emphasising the importance of accurate differentiation for effective therapeutic decision-making and enhanced patient outcomes.
4.1. Comparison with Related Methodological Literature
In this study a correlation of radiomic texture features extracted from computed tomography images and SNP-based microarray copy number variation cytogenomic features was performed. From the findings of the study, it is evident that there is a correlation between radiomic and genomic features, an outcome which has never been found by other studies in the distinction of chromophobe and oncocytoma renal masses. Nonetheless, the research found that the radiogenomics model for features with correlation > 0.55 gave an accuracy, sensitivity, Specificity, and AUC of 81.25, 75.00, 87.50 and 85.00 respectively.
The research had 16303 genes in total, out of which 97 genes were either significantly overlapping between the two tumour subtypes or were majorly present in either of the tumours. These genes were found in 61 different cytobands and were therefore correlated with tumour subtype. On correlation 24 cytobands containing 28 affected genes were found to be highly correlated to histopathology. These cytobands were found in chromosomes 1, 2, 6, 10, 17 and X. This is comparable to previous studies [
41,
53,
54,
55,
56,
57,
58,
59] which found these chromosomes to be associated with either chromophobe or oncocytoma. Specifically, 1p34.1 (RNF115), 1q21.3 (CTSK), 1q21.3 (S100A1), 1q22 (MUC1, RAB25), 1q25.2 (ANGPTL1), 1q32.3 (MTF2), 1q42.13 (TMED5), 1q21.2 (MCOLN2, MCOLN3), 1q32.1 (LAPTM5) and 1p36.22 (NBL1) were all found in chromosome 1.
RNF115 has emerged as a significant gene in our study, as well as in the broader study of renal tumours, particularly in differentiating between chromophobe renal cell carcinoma (ChRCC) and renal oncocytoma (RO). Research indicates that RNF115 is consistently expressed in all cases of renal oncocytoma and in oncocytic neoplasms favouring oncocytoma, but it is barely detectable in ChRCC [
33,
60,
61,
62,
63]. The study by Iakymenko et al. [
64] investigated the expression of the CTSK gene and its product, Cathepsin K, in RO and ChRCC. The findings revealed that Cathepsin K was positively expressed in both tumour types, with stronger staining observed in renal oncocytoma compared to the weaker, more membranous staining in ChRCC. This expression pattern suggests that while Cathepsin K is present in both types of tumours, the differences in staining intensity might serve as a definitive marker for differentiating between the two. In the study by Li et al. [
65], the S100A1 gene was expressed in 7 out of 8 RO cases, but not in any of the ChRCC cases. This gene expression pattern further supports the use of S100A1 as a diagnostic tool for differentiating between RO and ChRCC. These findings suggest that S100A1 is a useful marker, providing a reliable method for the differential diagnosis of renal RO and ChRCC. The study by Yusenko [
66] discussed the expression of the MUC1 gene in the context of differentiating ChRCC from RO. MUC1, also known as epithelial membrane antigen (EMA), showed higher expression levels in ChRCC compared to RO. The study highlighted that while MUC1 is expressed in both tumour types, its stronger and more consistent expression in ChRCC makes it a useful marker in the differential diagnosis between these two renal tumours [
66,
67]. The findings in our study align with previous research, suggesting that copy number variations (CNVs) in RNF115, CTSK, S100A1, MUC1, RAB25 [
68,
69], ANGPTL1 [
68], MTF2 [
70], TMED5 [
70], MCOLN2 [
68], MCOLN3 [
68], LAPTM5 [
68] and NBL1 [
71,
72] could be valuable biomarkers for distinguishing benign RO from malignant ChRCC, which is crucial for accurate diagnosis and treatment planning.
The present research found only a single cytoband in Chromosome 2 containing ERBB4 gene 2q24 (ERBB4). The study by Liu et al. [
73] demonstrated that hemizygous deletions of the ERBB4 gene were found in 33% of ChRCC cases, but not in any RO cases, indicating that ERBB4 deletions could serve as a useful marker for distinguishing between these two tumour types. In our study we found copy number alteration of 2q24 (ERBB4) in 38% of the ChRCC and none in RO which is comparable to Liu et al. [
73].
In a study aimed at distinguishing ChRCC from RO, the genes LMBRD1, TPBG, MANEA, and HACE1 were integral components of a 30-gene signature known as chromophobe and oncocytoma related gene signature (COGS). These genes were selected based on their differential expression patterns, which were identified through univariate gene expression and ROC curve analyses. The inclusion of these genes in the COGS signature contributed to the study’s ability to achieve a classification accuracy of 97.8% in the discovery dataset and 100% in the validation dataset, effectively differentiating ChRCC from RO using machine learning models. The cytobands 6q13 (LMBRD1), 6q14.1 (TPBG), 6q14.1 (MANEA), and 6q16.3 (HACE1) were found in our study to differentiate RO from ChRCC [
68].
According to the study done by Yusenko et al. [
70], the genes PRKG1 and CSTF2T are located within a region on chromosome 10q11.23-q21.1, where overlapping alterations were observed in both ChRCC and RO. The study by Krill-Burger et al. [
57] found that both ChRCC and RO exhibit significant genomic alterations, including copy number variations, in regions where the MRC1 and STAM genes are located. Specifically, deletions involving MRC1 and STAM were identified in ChRCC, with MRC1 being entirely deleted and STAM partially deleted. These deletions were significant in distinguishing ChRCC from other types, including RO. In our study 10q11.23 (PRKG1), 10q22.1 (CSTF2T), 10p12.33 (MRC1) and 10p12.1 (STAM) occurred in 37.5% in ChRCC and non in RO indicating the potential of the two cytoband in differentiating the two tumours.
In the study conducted by Satter et al. [
68], PPP3CB was identified as one of the top 197 genes through differential gene-expression and receiver-operating characteristic (ROC) analysis. This gene demonstrated a significant area under the curve (AUC) of 0.9 or higher, underscoring its potential role in distinguishing between chromophobe renal cell carcinoma and renal oncocytoma. Similarly, our study identified a copy number variation in the cytoband 10q22.2, which includes PPP3CB, in 33.33% of renal oncocytoma cases, with no such variation observed in chromophobe renal cell carcinoma cases.
The SLC4A1 gene, which encodes for a solute carrier family 4 member 1, plays a significant role in differentiating between RO and ChRCC. According to the study conducted by Molnar et al. [
74] SLC4A1 was expressed in 60% of RO but only in 11% of ChRCC. This difference in expression suggests that while SLC4A1 is more commonly associated with ROs, its lower expression in ChRCCs can still be present, albeit less frequently. The findings suggest that SLC4A1, could be used in the differential diagnosis between RO and ChRCC, especially when morphological features overlap [
74,
75]. In our study the cytoband 17q21.31 (SLC4A1) occurred in only in 16.67% of RO which does not provide sufficient proof on its ability to distinguish RO from ChRCC.
In the study by Satter et al. [
68] the DMD and DYNLT3 genes are included in the list of 197 top genes identified for their potential to differentiate ChRCC from RO. These genes were selected based on their differential expression and their ability to contribute to a gene signature (COGS) aimed at distinguishing between these two types of renal tumours. In our study Xp21.2 (DMD) and Xp11.23 (DYNLT3) occurred in 87.5% of ChRCC and 33.33% in RO patients.
The CTAG1B gene, also known as NY-ESO-1, was found to be expressed in 6 out of 18 ChRCCs and 15 out of 17 RO, suggesting its potential utility in diagnosing these tumours [
76]. The study by Demirović et al. [
77] investigated the expression of MAGE-A3/4 and NY-ESO-1 in RO and ChRCC, finding significant differences in the expression of these cancer testis antigens between the two tumour types, which may have diagnostic implications. In our study Xq28 (CTAG1B, MAGEA4, MAGEA3) occurred in 75% of ChRCC and 33.33% RO cases.
In radiomics, a total of 1,875 features [
25] were initially extracted, however after applying several feature reduction techniques, this number was reduced to 13 final features. These selected features belong to five radiomic feature classes and four filter classes. Among the final features, two were First Order features—Skewness and Minimum—each associated with three filter classes: ’Log Sigma 3 mm 3D’, ’Wavelet LLL’, and ’LBP 3D k’. The GLCM class contributed one feature, the ’Informational Measure of Correlation 2’ (IMC2), which was combined with the ’Wavelet LLH’ filter. Additionally, two GLDM features were selected: ’Large Dependence Low Gray Level Emphasis’ (LDLGLE) and ’Large Dependence High Gray Level Emphasis’ (LDHGLE), each combined with the ’Wavelet LLL’ and Logarithm filters. The GLRLM class included two features: ’Low Gray Level Run Emphasis’ (LGLRE) and ’Short Run Low Gray Level Emphasis’ (SRLGLE), which were combined with the ’Log Sigma 3 mm 3D’ and ’Log Sigma 2 mm 3D’ filters. Finally, one GLSZM feature having three filter types: ’Log Sigma 2 mm 3D’, ’Wavelet LHL’, and ’Wavelet LLH’ was selected.
The ’Log Sigma 3 mm 3D First Order Skewness’ and ’Wavelet LLL First Order Skewness’ are both radiomic features that measure the asymmetry of the intensity distribution within a 3D medical image [
25,
78], but they do so using different filtering techniques. The ’Log Sigma 3 mm 3D First Order Skewness’ involves applying a ’3D Gaussian’ smoothing filter with a ’sigma of 3 mm’, followed by a logarithmic transformation of the image intensities. This process enhances subtle textural details, particularly in lower intensity ranges, and the skewness metric quantifies the asymmetry in the distribution of these intensities. A positive Skewness indicates that the distribution leans towards lower intensity values, while a negative Skewness suggests a bias towards higher intensities. This property is particularly useful in highlighting variations in tissue composition that may be indicative of specific pathologies. On the other hand, the ’Wavelet LLL First Order Skewness’ is derived from a different type of filter—the wavelet transform. The ’Wavelet LLL’ filter applies low-pass filtering across all three dimensions (horizontal, vertical, and diagonal), which smoothens the image and emphasises large-scale, low-frequency components [
79]. After this transformation, the Skewness is calculated to assess the asymmetry of the intensity distribution in the filtered image. This feature is effective in capturing broader structural patterns within the tissue, which can be crucial for distinguishing between different types of tissues or abnormalities. In summary, while both features measure Skewness, the ’Log Sigma 3 mm 3D First Order Skewness’ focuses on fine details and intensity variations, particularly in lower intensity ranges, and the ’Wavelet LLL First Order Skewness’ emphasises larger structural patterns by smoothing the image across multiple dimensions. Both features provide complementary insights into the textural characteristics of tissues, aiding in the differentiation of complex medical conditions like Chromophobe Renal Cell Carcinoma and Renal Oncocytoma. Our findings are similar to what have been highlighted by previous research [
80,
81,
82,
83].
’LBP 3D k First Order Minimum’ is a combination of Local Binary Patterns (LBP) in three dimensions with ’First Order statistical Minimum’ value. LBP is a texture descriptor that captures the local spatial structure of images by analysing the relationship between a pixel and its surrounding neighbours [
84]. When applied in 3D, it extends this analysis to volumetric data, making it highly effective for capturing complex texture patterns in medical images [
85,
86,
87,
88,
89]. The ’First Order Minimum’ aspect focuses on the lowest intensity value in the voxel intensity distribution, providing insight into the darkest or least intense areas within the segmented volume [
90]. This combination is particularly useful in radiomics for identifying and characterising subtle variations in texture, which could be indicative of specific tissue properties or pathological conditions.
The radiomic feature ’Wavelet LLH GLCM IMC2’ represents a combination of wavelet transformation and Gray-Level Co-occurrence Matrix (GLCM) analysis focused on the ’Informational Measure of Correlation 2’ (IMC2) [
80,
82,
83,
91]. Wavelet transformation is a powerful tool that decomposes an image into different frequency components, allowing for the analysis of various levels of detail [
79]. The ’LLH filter’ specifically applies low-pass filtering in the first two dimensions (L and L) and high-pass filtering in the third dimension (H), capturing the horizontal details within the image. GLCM is a texture analysis method that evaluates the spatial relationship between pixel intensities, and IMC2 is a specific feature derived from GLCM, which quantifies the complexity of the texture by measuring the correlation between pixel pairs in the image. High values of IMC2 indicate a more complex and less predictable texture [
92,
93,
94]. By combining these techniques, the ’Wavelet LLH GLCM IMC2’ feature provides a sophisticated measure of texture that is sensitive to subtle patterns in the image, particularly those related to structural complexity and spatial relationships, making it valuable in distinguishing between ChRCC and RO.
The three radiomic features—’Wavelet LLL GLDM Large Dependence Low Gray Level Emphasis’, ’Logarithm GLDM Large Dependence Low Gray Level Emphasis’, and ’Logarithm GLDM Large Dependence High Gray Level Emphasis’—are advanced texture metrics used in radiomic analysis to capture subtle tissue characteristics in medical images [
25]. The Gray Level Dependence Matrix (GLDM) features focus on the relationship between a voxel and its dependent neighbours, emphasising different aspects of texture. ’Wavelet LLL GLDM Large Dependence Low Gray Level Emphasis’ is derived from applying a Wavelet transformation with a low-pass filter across all three axes (LLL), which highlights the broader, smooth patterns in the image. The ’Large Dependence Low Gray Level Emphasis’ then emphasises regions in the image where large groups of low-intensity pixels are clustered together, capturing homogeneity in low-density areas [
95]. ’Logarithm GLDM Large Dependence Low Gray Level Emphasis’, is similar to the first but uses a logarithmic transformation instead of a wavelet filter. The logarithm filter can enhance subtle differences in pixel intensity, making this feature particularly useful for detecting fine, low-intensity patterns in the image that might be missed by other filters. ’Logarithm GLDM Large Dependence High Gray Level Emphasis’, unlike the previous two, emphasises areas with large clusters of high-intensity pixels. The logarithmic transformation again helps to enhance the contrast and detail within these high-intensity regions, making this feature useful for identifying dense or bright areas within the image that may correlate with certain pathological features [
96,
97]. Together, these features allow for a detailed analysis of the image’s texture, capturing both low- and high-intensity patterns that can be crucial for distinguishing between different tissue types or identifying specific pathological changes.
The three radiomic features—’Log Sigma 3 mm 3D GLRLM Low Gray Level Run Emphasis’, ’Log Sigma 2 mm 3D GLRLM Short Run Low Gray Level Emphasis’, and ’Log Sigma 3 mm 3D GLRLM Short Run Low Gray Level Emphasis’—are texture measures derived from the Gray-Level Run Length Matrix (GLRLM) combined with specific logarithmic filters applied to 3D images [
25]. ’Log Sigma 3 mm 3D GLRLM Low Gray Level Run Emphasis’ focuses on the emphasis of runs of low gray-level values, highlighting regions with low-intensity pixels that are clustered together [
98]. The ’Log Sigma 3 mm 3D’ filter applied to this feature enhances finer details within the image at a specific spatial scale, making it useful for identifying subtle low-intensity structures within the volume. ’Log Sigma 2 mm 3D GLRLM Short Run Low Gray Level Emphasis’ measures the emphasis on shorter runs of low-intensity pixels, which indicates a texture where these pixels appear in smaller, more isolated clusters. The ’Log Sigma 2 mm 3D’ filter is used here to capture finer, more localised texture patterns, emphasising the presence of smaller-scale low-intensity areas in the image. ’Log Sigma 3 mm 3D GLRLM Short Run Low Gray Level Emphasis’ similar to the second feature, also emphasises short runs of low gray-level pixels but with a ’Log Sigma 3 mm 3D filter’. This filter size captures slightly larger texture patterns compared to the 2 mm filter, allowing the feature to identify small but slightly broader low-intensity areas, which could be indicative of certain pathological changes or tissue characteristics. Together, these features provide a nuanced analysis of the texture in medical images, particularly focusing on low-intensity regions, which can be critical for detecting and characterising specific tissue properties or abnormalities [
99,
100].
The three radiomic features—’Log Sigma 2 mm 3D GLSZM Small Area Low Gray Level Emphasis’, ’Wavelet LHL GLSZM Small Area Low Gray Level Emphasis’, and ’Wavelet LLH GLSZM Small Area Low Gray Level Emphasis’—are derived from the Gray-Level Size Zone Matrix (GLSZM), a texture analysis method that quantifies the size of homogeneous zones of gray levels in an image, combined with specific filters that enhance different aspects of the image texture [
25,
98]. ’Log Sigma 2 mm 3D GLSZM Small Area Low Gray Level Emphasis’ emphasises small areas within the image that consist of low gray-level zones, highlighting regions where small clusters of low-intensity pixels are prevalent. The ’Log Sigma 2 mm 3D’ filter enhances the detection of fine texture details at a specific spatial scale, making this feature useful for identifying subtle patterns of low-intensity areas in the image. ’Wavelet LHL GLSZM Small Area Low Gray Level Emphasis’ [
25] is a feature in which the ’Wavelet LHL’ filter is applied capturing the horizontal high-frequency details along with low-pass filtering in the other directions. This combination focuses on small, low-intensity zones in the image, particularly those with fine horizontal structures, allowing for detailed texture analysis in specific directions. ’Wavelet LLH GLSZM Small Area Low Gray Level Emphasis’ is similar to the second feature, it applies the ’Wavelet LLH’ filter, which emphasises high-frequency details in the vertical direction while applying low-pass filtering horizontally. This feature targets small areas of low-intensity zones, especially those aligned with vertical structures, providing a focused analysis of these specific patterns within the image. These features collectively contribute to a detailed texture analysis by focusing on small, low-intensity areas within the image, enhanced by various filters that capture specific directional details. This facilitates the recognition of fine texture details that could be pivotal in distinguishing various tissue types or pinpointing specific pathological changes in medical imaging [
101].
It’s worth noting that the radiomic features extracted from the 14 patients did not achieved statistical significance. However, a previous study by Alhussaini et al. [
22], involving a larger cohort of 78 patients found that at least four features either attained or approached statistical significance. This finding highlights the vital role that sample size plays in enhancing the statistical power of analyses, demonstrating how a larger sample can reveal significant trends that smaller samples may not capture.
In conclusion, our research identified significant correlations between specific radiomics features and genomic markers, highlighting the potential of radiogenomics in non-invasive tumour characterisation. Notably, ’Log Sigma 3 mm 3D Firstorder Skewness’ showed strong correlations with ChXp21.2 (DMD) (-0.73), ChXp11.23 (DYNLT3) (-0.73), and Ch2q24 (ERBB4) (-0.65). Additionally, ’Logarithm GLDM Large Dependence High Gray Level Emphasis’ was linked with Ch6q14.1 (TPBG) (-0.61), while ’Wavelet LLL Firstorder Skewness’ correlated with Ch6q14.1 (TPBG) (-0.61), Ch6q13 (LMBRD1) (-0.58), Ch6q14.1 (MANEA) (-0.58), and Ch6q16.3 (HACE1) (-0.58). Finally, ’Wavelet LHL GLSZM Small Area Low Gray Level Emphasis’ was associated with ChXp21.2 (DMD) (-0.57) and ChXp11.23 (DYNLT3) (-0.57). These findings underscore the potential of radiomics features as surrogates for genomic data, offering promising avenues for enhancing non-invasive diagnostic and prognostic tools in clinical practice.