Predicting high proliferative Ki-67 index in patients with adrenocortical carcinoma based on texture analysis of contrast-enhanced computed tomography images: a cross-sectional study

Full Text

BACKGROUND

Adrenocortical carcinoma (ACC) is a rare malignant tumor of the adrenal cortex, typically characterized by an aggressive clinical course and an unfavorable prognosis [1]. The overall 5-year survival is 60%–80% for tumors confined to the adrenal gland, 35%–50% for locally advanced disease, and less than 28% for metastatic disease [2]. Rare cases of a more favorable, indolent disease course with slow progression and late metastasis have been described. The only potentially curative treatment option for ACC is timely surgical resection of the tumor. The urgency of surgical intervention is associated with rapid tumor growth and the high likelihood of metastasis [1, 2].

ACC is characterized by marked variability in morphological features (classical, oncocytic, myxoid, and sarcomatoid variants), proliferative activity, clinical course, response to therapy, as well as in overall and recurrence-free survival associated with these factors [1–3]. The heterogeneity of ACC underlies the difficulty of predicting the disease course and selecting optimal therapeutic strategies [4]. One of the key predictors of ACC recurrence is the Ki-67 proliferation index. At Ki-67 values > 10%, the probability of ACC recurrence, even after adequate resection with negative margins (R0 resection), is at least 80%. With Ki-67 ≤10%, the prognosis is more favorable, and adjuvant chemotherapy is not indicated [1]. Ahmed et al. [5] evaluated the correlation between the Ki-67 proliferation index and texture features of contrast-enhanced computed tomography (CT) images in 53 patients with ACC. The authors demonstrated that the tumor geometric features of elongation and flatness discriminate ACC cases with high (Ki-67 > 10%) and low (Ki-67 ≤10%) proliferation indices, with areas under the receiver operating characteristic curve (AUC-ROC) of 0.70 and 0.78, respectively. In addition, the multivariate linear regression model yielded the following parameters: coefficient of determination R² = 0.67, adjusted R² = 0.462, and Pearson correlation coefficient r = 0.824. Significant features included shape elongation, shape flatness, gray level co-occurrence matrix (GLCM) cluster shade, gray level run length matrix (GLRLM) long run emphasis, interquartile range of attenuation values distribution, neighboring gray tone difference matrix (NGTDM), and gray level size zone matrix (GLSZM). Several other studies have examined the accuracy of texture analysis and deep learning methods for the binary classification of adrenal lesions of indeterminate CT phenotype according to malignancy criteria [6–8]. However, the results of these studies remain preliminary, and the work by Tucci et al. [8], which employed a neural network model and radiomics, was published only as conference proceedings. Therefore, the model architecture, its parameters, and the characteristics of the study cohort remain unclear. Nevertheless, contrast-enhanced CT, owing to its higher spatial resolution compared with other imaging modalities, represents an accessible and informative method for the detection of ACC [1, 2].

AIM

To evaluate the accuracy of predicting a high Ki-67 proliferation index in patients with ACC using radiomics-based texture analysis of contrast-enhanced CT images.

METHODS

Study Design

A single-center, non-comparative, cross-sectional, retrospective study was conducted with respect to CT acquisition, and a prospective study with respect to repeat analysis of CT images.

Study Setting

The study included data from patients registered in the medical information system of the National Medical Research Center for Endocrinology named after Academician I.I. Dedov (Moscow) between April 2014 and October 2024. To identify eligible cases within the information system, an initial search was performed to compile a list of all patients who had undergone surgical treatment for a malignant neoplasm of the adrenal gland (code C74.0 of the International Classification of Diseases, 10th Revision) in the Surgical Department of the National Medical Research Center for Endocrinology. During the study period, CT examinations were performed in the Department of Computed Tomography and Magnetic Resonance Imaging, and immunohistochemical examinations were conducted in the Department of Pathomorphology of the same center.

Eligibility Criteria

Inclusion criteria:

• Histologically confirmed diagnosis of ACC;
• Availability of preoperative four-phase contrast-enhanced CT images of the abdominal organs obtained no more than 1 month prior to surgical intervention (non-contrast, arterial, venous, and delayed phases).

Exclusion criteria: presence of artifacts in the adrenal region on CT images (motion artifacts, ring artifacts).

Computed Tomography Acquisition

The study included images obtained using the following CT scanners:

• Optima® CT660 (GE HealthCare, USA) in use from 2015 to the present;
• Revolution® CT (GE HealthCare, USA) in use from 2019 to the present; and
• BrightSpeed® 16 (GE HealthCare, USA) used in 2014.

The CT scanner and image reconstruction parameters are presented in Table 1. Medrad Stellant® (Bayer, Germany) with a dual-syringe automatic injector at an injection rate of 3.5–4 mL/s was used for contrast enhancement throughout the data acquisition period. The arterial phase was acquired 10 s after triggering of the bolus tracker positioned in the descending aorta at the level of the diaphragm (120 HU), the venous phase at 30 s after bolus triggering, and the delayed phase at 10–15 min after contrast agent administration.

 

Table 1. Parameters of computed tomography and image reconstruction
Parameter	Scanning phase
	Non-contrast	Arterial	Venous	Delayed
Optima® CT660 (GE HealthCare, USA)
Tube voltage, kV	120, 100	120, 100	120, 100	120, 100
Tube current, mA	automatic modulation 380, 390, 410, 480	automatic modulation 380, 400	automatic modulation 335, 390, 400	automatic modulation 250
Exposure, mA×s	9, 14, 18, 19, 20, 22, 24, 25, 26, 28	4, 5, 6, 7, 8, 9	3, 5, 6, 7, 8, 9	3, 5, 7, 8, 9, 12, 20
Slice spacing, mm	0.625	0.625	0.625	0.625
Slice thickness, mm	1.25	1.25	1.25	1.25
Reconstruction kernel	Standard	Standard	Standard	Standard
Revolution® CT (GE HealthCare, USA)
Tube voltage, kV	100, 120	100, 140	140	140
Tube current, mA	automatic modulation 100, 120	automatic modulation 100, 240, 405, 485	automatic modulation 240, 405, 485	automatic modulation 240, 320, 485
Exposure, mA×s	1, 2, 3	1, 2, 4, 5	2, 4, 9, 10	2, 3, 4, 6
Slice spacing, mm	0.625	0.625 и 1.25	1.25	1.25
Slice thickness, mm	0.625	0.625 и 1.25	1.25	1.25
Reconstruction kernel	Standard	Standard	Standard	Standard
BrightSpeed® 16 (GE HealthCare, USA)
Tube voltage, kV	120	120	120	120
Tube current, mA	400	380	380	380
Exposure, mA×s	28	27	27	27
Slice spacing, mm	1.25	1.25	1.25	1.25
Slice thickness, mm	1.25	1.25	1.25	1.25
Reconstruction kernel	Standard	Standard	Standard	Standard

 

Computed Tomography Image Analysis

Image Segmentation

Preoperative CT images were exported from the Picture Archiving and Communication System (PACS) to a personal computer in DICOM (Digital Imaging and Communications in Medicine) format. Prior to processing, all data were anonymized, i.e., any information enabling patient identification such as name, date of birth, identification number, or date of examination, was removed. Regions corresponding to adrenal lesions were manually segmented by radiologists (with more than 5 years of professional experience) using the 3D Slicer® software package, version 5.6.2 (Slicer Community, USA), separately for each scanning phase. Segmentation files were saved in DICOM format together with the original CT images. Re-analysis was blind, without access of the radiologists to clinical information, including previous CT findings and laboratory data (except for the most recent data obtained before surgery).

Texture Analysis of Images

CT images together with segmentation files were exported to the PyRadiomics® platform, version 3.1.0 (Computational Imaging & Bioinformatics Lab, USA). No additional filters were used for image analysis. After resampling to an isotropic voxel size of 1 × 1 × 1 mm in 3D, 106 features belonging to seven feature classes were determined for each image (each CT phase):

• First order statistics: 18 features;
• Shape-based: 14 features;
• Gray Level Co-occurrence Matrix (GLCM): 23 features;
• Gray Level Run Length Matrix (GLRLM): 16 features;
• Neighboring Gray Tone Difference Matrix (NGTDM): 5 features;
• Gray Level Dependence Matrix (GLDM): 14 features; and
• Gray Level Size Zone Matrix (GLSZM): 16 features.

Feature extraction was performed within a Hounsfield unit window corresponding to the abdominal imaging protocol [−160; 240] HU. All other image preprocessing parameters (normalization, interpolation algorithm, presence of threshold-based segmentation for region-of-interest refinement, etc.) were left at their default values in accordance with the PyRadiomics module settings [9].

Predicted Outcome

A > 10% Ki-67 proliferation index was used to assess the predictive accuracy of CT image texture analysis. The 10% threshold was selected based on its clinical significance and its use for discriminating tumors with low and high proliferative activity, which determines prognosis and therapeutic strategy in patients with ACC [1].

Sections stained with hematoxylin and eosin using standard techniques were prepared from histologic ACC specimens obtained during surgical treatment at the National Medical Research Center for Endocrinology. Standard adrenal tissue processing was performed using a Leica® ASP6025 S tissue processor (Leica Biosystems, Germany), followed by paraffin embedding.

Midline sections were prepared using a Leica® RM 2125 RTS microtome (Leica Biosystems, Germany) and, after deparaffinization, stained with hematoxylin and eosin on a Leica® ST5010 AXL stainer (Leica Biosystems, Germany).

Histologic evaluation of adrenal pathology slides was performed by light microscopy using a Leica® DM2500 microscope (Leica Microsystems, Germany) and computerized morphometry with Aperio® ImageScope software (Leica Microsystems, Germany).

All adrenocortical tumors were assessed by the following pathological criteria:

• Tumor size and weight;
• Nuclear polymorphism;
• Mitotic count in 10 high-power fields at ×400 magnification;
• Tumor cell cytoplasm (0%–25% vs 26%–100% clear) and growth pattern (diffuse vs non-diffuse); and
• Presence or absence of atypical mitoses, necrosis, and definite capsular, venous, sinusoidal, and adjacent organ invasion.

Tumor malignant potential was evaluated using the Weiss score. In the case of oncocytic adrenal neoplasms, characterized by granular and intensely eosinophilic cytoplasm, marked nuclear polymorphism, and diffuse growth pattern, the modified scoring Lin–Weiss–Bisceglia system was applied. All adrenocortical carcinomas in this study met ≥4 Weiss histological criteria [10].

Mitotic rate was determined by counting 50 high-power fields (×400) using the Leica® DM2500 microscope (Leica Microsystems, Germany). Selected areas of the slides contained the highest density of mitotic figures. Each set of 10 high-power fields was counted on different slides whenever possible.

Tumor architecture was defined as diffuse when more than 33% of its cell layers lacked a characteristic growth pattern.

Vessels lined by endothelium and incorporating a muscular wall were considered veins, whereas sinusoids were defined as vessels with an endothelial lining and minimal supporting tissue. Venous or sinusoidal invasion was recorded when tumor cells were identified within the lumen of such vessels adjacent to the vessel wall, both inside and outside the adrenal tumor. Capsular invasion was defined as complete tumor penetration through the surrounding capsule.

Immunohistochemical assay was performed on a fully automated Leica® Bond III staining system (Leica Microsystems, Germany) according to the standard protocols recommended by the manufacturer. Anti–Ki-67 monoclonal antibodies (MIB-1 at 1:150 dilution; Dako, Denmark) were used for the immunohistochemical assay. Tumor cell proliferative activity (Ki-67) was assessed visually in 10 high-power fields at ×400 magnification. The value was determined as the percentage of positively stained tumor cell nuclei in areas with the highest labeling activity (hot spots). The immunohistochemical methodology remained unchanged throughout the study period.

Group Analysis

The patients were divided into two groups based on the Ki-67 index value:

• Group 1: Ki-67 ≤10%; or
• Group 2: Ki-67 > 10%.

Ethics Approval

The study protocol was approved by the Local Ethics Committee of the National Medical Research Center for Endocrinology named after Academician I.I. Dedov (Minutes No. 20 dated November 13, 2024). All patients, when seeking medical care at the Center, provided informed consent for the use of the results of examinations and treatment for scientific purposes.

Statistical Analysis

Sample size calculation: The required sample size was not calculated at the study planning stage.

Missing data: When forming the study sample, it was found that all patients included in the analysis after confirmed compliance with the inclusion and exclusion criteria had a complete set of required data. No cases of missing data were identified.

Statistical methods: Python, version 3.9.21, was used for data analysis.

Data for continuous variables (age at the time of contrast-enhanced CT [years]; maximum linear size of the adrenal lesion on CT [mm]; CT attenuation values by phase in Hounsfield units [HU]; Ki-67 index value [%]) are presented as M ± SD, where M is arithmetic mean and SD is standard deviation. For comparison of patient groups by categorical variables, Fisher’s exact test was used; for continuous variables, the Mann–Whitney U test was applied.

Post-processing of texture features included standardization and selection of the most informative features. Standardization was performed by sample mean centering (subtracting the overall sample mean value of each feature from each feature value) and scaling by the standard deviation (dividing the result of mean subtraction for each feature value by the overall sample standard deviation). Prior to selecting the most significant features, normality of distribution was assessed using the Shapiro–Wilk test. When p <  0.05, the two-sided Mann–Whitney U test was used to compare groups 1 and 2. When the Shapiro–Wilk test did not reveal significant deviation from normality (p ≥ 0.05), the two-sided Student t-test was applied. Features demonstrating significant between-group differences were used for cluster analysis.

The applicability of CT texture features for differentiating between groups 1 and 2 was evaluated using cluster analysis with the fuzzy K-means algorithm. This method allows assessment of whether the data naturally separate into groups and whether an internal structure exists corresponding to biologically meaningful subgroups [11]. The fuzzy K-means clustering method was chosen for its ability to determine the probability of membership in a given cluster for each point in principal component space. The feature space for clustering was formed after dimensionality reduction using principal component analysis, which was also applied to eliminate correlated features and to extract the most informative components. The number of principal components was determined based on the proportion of explained variance and the feasibility of visual interpretation, retaining components that together explained at least 90% of the total variance. The cluster analysis included the following steps:

• Formation of the feature matrix after preprocessing;
• Application of the fuzzy K-means algorithm with iterative refinement of cluster centers; and
• Evaluation of the resulting distribution of observations across clusters and interpretation of the identified groups in terms of their correspondence to the original classification by Ki-67 index value.

The analysis was based on the assumption that the principal patterns in the data can be expressed as linear combinations of the original features. Clustering was performed under the assumption that differences in CT image features may reflect boundaries between groups with different levels of tumor proliferative activity.

RESULTS

Study Sample Formation

Between 2014 and 2024, surgical removal of malignant adrenal tumors was performed in 51 patients in the Department of Surgery of the National Medical Research Center for Endocrinology. Pathological verification of the diagnosis was obtained in 42 cases, among which contrast-enhanced CT images were available for 31 patients. One patient was excluded due to pronounced adrenal region artifacts on CT images, and six patients were excluded because of the absence of immunohistochemical study results. Consequently, data from 24 patients with ACC were included in the final analysis:

• Group 1: Ki-67 index ≤ 10% (n = 9);
• Group 2: Ki-67 index > 10% (n = 15).

Characteristics of Study Groups

According to contrast-enhanced CT performed using Optima® CT660 (GE HealthCare, USA) in 10 patients, Revolution® CT (GE HealthCare, USA) in 13 patients, and BrightSpeed® 16 (GE HealthCare, USA) in 1 patient, the maximum contrast enhancement was observed in the venous phase, followed by moderate washout in the delayed phase (Table 2). Patients in groups 1 and 2 were comparable with respect to sex, age, and most adrenal lesion parameters, except for venous-phase attenuation, which was significantly higher in group 1 (Table 2).

 

Table 2. Comparative characteristics of patients in the groups
Parameter	Group 1, n = 9	Group 2, n = 15	р
Male sex, n (%)	4 (44)	4 (27)	0.412
Age, years	41.4 ± 15.2	47.1 ± 14.3	0.220
Adrenal lesion characteristics
Maximum linear size, mm	73.9 ± 29.8	92.4 ± 28.9	0.152
Attenuation in non-contrast phase, HU	34.6 ± 5.4	34.5 ± 5.5	0.787
Attenuation in arterial phase, HU	63.6 ± 16.9	53.5 ± 11.9	0.160
Attenuation in venous phase, HU	93.7 ± 26.7	68.9 ± 13.2	0.027
Attenuation in delayed phase, HU	57.9 ± 10.6	53.7 ± 4.5	0.339

 

Primary Results

With the Shapiro–Wilk test, the data of 71 features (67%) extracted from CT images of patients with ACC followed a normal distribution. These features were therefore compared between groups using the Student t-test. The Mann–Whitney U test was applied in all remaining cases. The features significant for the purposes of this study are presented in Table 3. The venous phase of contrast-enhanced CT proved to be the most informative for quantitative analysis. Subsequent dimensionality reduction of the feature space allowed the identification of three principal components, which together explained 99.6% of the total variance in the dataset.

 

Table 3. Significant computed tomography texture features
Computed tomography phase	Significant features
Non-contrast	None
Arterial	None
Venous	firstorder_10Percentile, firstorder_90Percentile, firstorder_Mean firstorder_Median, firstorder_RootMeanSquared, glcm_Autocorrelation, glcm_SumAverage, gldm_HighGrayLevelEmphasis, gldm_LargeDependenceLowGrayLevelEmphasis, gldm_LowGrayLevelEmphasis, gldm_SmallDependenceHighGrayLevelEmphasis, glrlm_HighGrayLevelRunEmphasis, glrlm_LongRunHighGrayLevelEmphasis, glrlm_LongRunLowGrayLevelEmphasis, glrlm_LowGrayLevelRunEmphasis, glrlm_ShortRunHighGrayLevelEmphasis, glrlm_ShortRunLowGrayLevelEmphasis, glszm_HighGrayLevelZoneEmphasis, glszm_LargeAreaLowGrayLevelEmphasis, glszm_LowGrayLevelZoneEmphasis, glszm_SmallAreaHighGrayLevelEmphasis, glszm_SmallAreaLowGrayLevelEmphasis
Delayed	None
Note. With the Student t-test / Mann–Whitney U test, p <  0.05. Feature names reflect a structured hierarchy of parameters; for example, a feature named glszm_SmallAreaLowGrayLevelEmphasis consists of several components, each of which has a specific meaning: glszm indicates that the feature is based on GLSZM, and SmallAreaLowGrayLevelEmphasis is the specific feature name.

 

To improve interpretability of the analyzed data, Table 4 provides a detailed description of the selected features grouped by category. Description indicating which specific aspects of tumor texture organization they represent on CT images.

 

Table 4. Significant texture features and their interpretation
Feature group	Feature names	Interpretation
First-order features	• firstorder_10Percentile; • firstorder_90Percentile; • firstorder_Mean; • firstorder_Median; • firstorder_RootMeanSquared	Describe the distribution of voxel attenuation values within the region of interest without considering their spatial arrangement, corresponding to a qualitative characteristic such as mean attenuation
Gray level co-occurrence matrix–based features	• glcm_Autocorrelation; • glcm_SumAverage	Quantitatively assess spatial tissue heterogeneity by considering the relative positions of voxels, corresponding to qualitative characteristics such as homogeneous structure (increase in glcm_Autocorrelation) and predominance of pairs of hyperdense voxels (increase in glcm_SumAverage)
Gray level dependence matrix–based features	• gldm_HighGrayLevelEmphasis; • gldm_LargeDependenceLowGrayLevelEmphasis; • gldm_LowGrayLevelEmphasis; • gldm_SmallDependenceHighGrayLevelEmphasis	Quantitatively assess the size and attenuation of homogeneous zones in the image texture, corresponding to qualitative characteristics of hypodense regions (increase in gldm_LargeDependenceLowGrayLevelEmphasis, gldm_LowGrayLevelEmphasis) and hyperdense inclusions (increase in gldm_SmallDependenceHighGrayLevelEmphasis, gldm_HighGrayLevelEmphasis)
Gray level run length matrix–based features	• glrlm_HighGrayLevelRunEmphasis; • glrlm_LongRunHighGrayLevelEmphasis; • glrlm_LongRunLowGrayLevelEmphasis; • glrlm_LowGrayLevelRunEmphasis; • glrlm_ShortRunHighGrayLevelEmphasis; • glrlm_ShortRunLowGrayLevelEmphasis	Quantitatively assess the length and attenuation of linear homogeneous regions (runs of consecutive voxels with the same attenuation) in the image texture, corresponding to qualitative characteristics of homogeneous hyperdense regions (increase in glrlm_HighGrayLevelRunEmphasis, glrlm_LongRunHighGrayLevelEmphasis), homogeneous hypodense regions (increase in glrlm_LongRunLowGrayLevelEmphasis, glrlm_LowGrayLevelRunEmphasis), small hyperdense inclusions (increase in glrlm_ShortRunHighGrayLevelEmphasis), and small hypodense inclusions (increase in glrlm_ShortRunLowGrayLevelEmphasis)
Gray level size zone matrix–based features	• glszm_HighGrayLevelZoneEmphasis; • glszm_LargeAreaLowGrayLevelEmphasis; • glszm_LowGrayLevelZoneEmphasis; • glszm_SmallAreaHighGrayLevelEmphasis; • glszm_SmallAreaLowGrayLevelEmphasis	Quantitatively assess the size and intensity of connected homogeneous zones in the image texture, corresponding to qualitative characteristics such as homogeneous hyperdense regions (increase in glszm_HighGrayLevelZoneEmphasis), large homogeneous hypodense regions (increase in glszm_LargeAreaLowGrayLevelEmphasis, glszm_LowGrayLevelZoneEmphasis), small hypodense inclusions (increase in glszm_SmallAreaLowGrayLevelEmphasis), and small hyperdense inclusions (increase in glszm_SmallAreaHighGrayLevelEmphasis)

 

The clustering outcome within the resulting feature space is illustrated in Figure 1. Thus, two clusters form relatively isolated groups of points. The original distribution of cases according to the Ki-67 index is shown in Figure 2. The allocation of patients to clusters is summarized in Table 5: the first cluster predominantly includes patients from group 1, whereas the second cluster is mainly composed of patients from group 2. According to Fisher’s exact test, the resulting p value confirms that the fuzzy K-means algorithm partitioned the cohort into two clusters significantly associated with Ki-67 status.

 

Fig. 1. Visualization of the fuzzy K-means clustering results in the principal component space.

 

Fig. 2. Visualization of the initial distribution of adrenocortical carcinoma cases into groups according to the Ki-67 index value in the principal component space. Ki-67, proliferative activity index.

 

Table 5. Distribution of patients with adrenocortical carcinoma by clusters
Clusters	Group 1, n = 9	Group 2, n = 15	р
Cluster 1	5	1	0.015
Cluster 2	4	14
Note. The p value was calculated using Fisher’s exact test.

 

The heatmap demonstrates differences in mean feature values between the clusters (Fig. 3). It was found that, in the cluster with high Ki-67 values, the intensity of texture elements of ACC lesions on venous-phase CT images was on average lower than in the cluster with low Ki-67 values, as reflected by the first-order statistical features. Moreover, analysis of matrix-based texture features revealed a lower degree of texture homogeneity in the high Ki-67 cluster compared with the low Ki-67 cluster.

 

Fig. 3. Heat map of feature values in the groups. Color intensity ranges from 0 to 1, where 0 corresponds to the cluster with high Ki-67 index values and 1 corresponds to the cluster with low Ki-67 index values.

 

DISCUSSION

Summary of Primary Results

The fuzzy K-means algorithm reliably separates patients with ACC into two clusters associated with low (Ki-67 ≤10%) and high (Ki-67 > 10%) proliferative index values. The cluster with a high Ki-67 index is characterized by lower intensity and reduced texture homogeneity on venous-phase CT images compared with the cluster with low values.

Interpretation of Primary Results

The obtained results confirm the hypothesis that texture features extracted from contrast-enhanced CT images can be used to assess the Ki-67 index. This conclusion is consistent with the findings of other studies. For example, the study by Ahmed et al. [5] demonstrated that certain radiomic features may be associated with the level of proliferative activity of adrenocortical carcinoma as assessed by the Ki-67 index. Their multivariate linear regression model included such informative features as shape elongation, shape flatness, GLCM cluster shade, GLRLM long run emphasis, interquartile range (first order), NGTDM Contrast, and GLSZM Gray Level Non-Uniformity Normalized. In our study, not only the above-mentioned features but also additional ones were found to be significant, including firstorder_10Percentile, firstorder_90Percentile, firstorder_Mean, firstorder_Median, firstorder_RootMeanSquared, glcm_Autocorrelation, glcm_SumAverage, as well as texture parameters of GLDM and GLRLM. Nevertheless, Ahmed et al. [5] did not describe the details of image post-processing and the selection of radiomic feature extraction parameters; therefore, it can be assumed that the differences between the studies are due to variations in image processing methods and in the selection of preprocessing parameters. In addition, the characteristics of the study sample should be considered: the mean size of ACC in our patients was smaller (73.9 ± 15.2 mm and 92.4 ± 14.3 mm in group 1 and group 2, respectively) compared with the study by Ahmed et al. [5], where the mean lesion size was 115 ± 64.8 mm, which could have influenced the observed associations.

Furthermore, the results of our study are consistent with the data reported by Liu et al. [12], as both studies demonstrate the possibility of using features extracted from contrast-enhanced CT images as prognostically substantial parameters in ACC. The study by Liu et al. [12] showed that radiomics outperforms the conventional ENSAT and S-GRAS systems in predicting overall survival and recurrence-free survival. However, unlike that study, our work focuses on cluster analysis and the association between clusters and the Ki-67 index. Our results complement the existing body of scientific evidence by confirming that radiomics, in combination with the fuzzy K-means clustering algorithm, allows differentiation of ACC cases according to the Ki-67 index value, which is one of the key markers of malignancy [1].

According to our findings, the cluster with high Ki-67 index values is characterized by a lower degree of texture homogeneity than the cluster with low values. This observation is consistent with the results of the study by Robertson-Tessi et al. [13], in which modeling and experimental data demonstrated that tumors with pronounced heterogeneity exhibit more aggressive behavior, which may explain our finding of texture heterogeneity in the cluster with high Ki-67 values, also associated with tumor aggressiveness.

We also found that only the venous-phase images obtained by contrast-enhanced CT contained features considerable for the purposes of this study. In addition, in the cluster with high Ki-67 index values, the gray-level intensity on CT images was, on average, lower than in the cluster with low values, which may indicate reduced contrast accumulation in tumors with a high Ki-67 index.

Study Limitations

The use of the Ki-67 index as a marker of ACC progression risk has certain limitations. Although it is widely applied in routine clinical practice and considered an important prognostic indicator, it is not the only marker reflecting the biological aggressiveness of the tumor. The incorporation of additional information, such as molecular data or other adrenal tumor markers, could provide a more comprehensive understanding of the relationship between texture features extracted from contrast-enhanced CT images and tumor malignancy grade. For example, a high expression level of steroidogenic factor 1 (SF-1) is known to be a prognostic factor associated with unfavorable clinical outcomes in ACC [14], whereas insulin-like growth factor 2 (IGF-2) expression has been associated with longer overall survival [15]. Consideration of such markers in future studies may contribute to a more accurate interpretation of radiomic data and enhance the diagnostic value of CT-based models.

Another limitation of the present study is the relatively small sample size, which, among other things, may affect the generalizability (extrapolation beyond the study) of the obtained results. The inclusion of additional data from multidisciplinary medical centers and using different scanning protocols could improve model robustness and accuracy. A similar approach was used in the study presented at the 26th European Congress of Endocrinology [5], where the analysis of data from seven centers substantially increased the diagnostic accuracy of machine learning models. Higher predictive performance of radiomics may be achieved by employing more complex neural network architectures capable of capturing non-linear relationships between features. For example, Ahmed et al. [5] developed a neural network algorithm for binary classification based on malignancy criteria, achieving the following performance metrics: AUC-ROC of 0.974, sensitivity of 92.7%, and specificity of 92.8%.

The application of contrast-enhanced CT in our study was limited to cases of ACC and to a specific prognostic marker—the Ki-67 index. The use of this method for assessing the aggressiveness of other tumors or for analyzing other markers requires additional investigation.

CONCLUSION

The role of quantitative analysis of contrast-enhanced CT images in the non-invasive assessment of ACC progression risk has remained insufficiently studied to date: no publications on this topic have been available in Russia, and only isolated studies have been reported internationally. In the present study, an approach to the quantitative characterization of CT images in patients with ACC using fuzzy K-means clustering is proposed, allowing the identification of patients with potentially aggressive tumor behavior. An association between quantitative CT image features and tumor proliferative activity was established, and significant differences between clusters confirm the applicability of this method for assessing the risk of ACC progression. The obtained results demonstrate the potential of the developed approach for the preliminary estimation of the Ki-67 index based on texture features of contrast-enhanced CT. Further studies should focus on methodology standardization, integration of radiomics with clinical characteristics, and improvement of the reproducibility of this approach.

ADDITIONAL INFORMATION

Author contributions: A.V. Manaev: conceptualization, data curation, formal analysis, writing—original draft; N.V. Tarbaeva: conceptualization, project administration, writing—original draft; A.A. Roslyakova, D.G. Beltsevich, L.S. Urusova: data curation, investigation, resources, writing—review & editing; N.G. Mokrysheva, V.E. Sinitsyn: project administration, supervision, writing—review & editing. All the authors approved the version of the manuscript to be published and agreed to be accountable for all aspects of the work, ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Ethics approval: The study protocol was approved by the local Ethics Committee of the Dedov National Medical Research Center for Endocrinology (Minutes No. 20 dated November 13, 2024). All the patients provided written informed consent for the use of their clinical assessment and treatment data for research purposes on admission.

Funding sources: No funding.

Disclosure of interests: The authors have no relationships, activities, or interests for the last three years related to for-profit or not-for-profit third parties whose interests may be affected by the content of the article.

Statement of originality: No previously obtained or published material (text, images, or data) was used in this study or article.

Data availability statement: The editorial policy regarding data sharing does not apply to this work.

Generative AI: No generative artificial intelligence technologies were used to prepare this article.

Provenance and peer-review: This article was submitted unsolicited and reviewed following the standard procedure. The peer review process involved two external reviewers, a member of the Editorial Board, and the in-house science editor.

About the authors

Almaz V. Manaev

Endocrinology Research Centre; National Research Nuclear University “MEPhI”

Author for correspondence.
Email: a.manaew2016@yandex.ru
ORCID iD: 0009-0003-8035-676X
SPIN-code: 2902-9767

MD

Russian Federation, Moscow; Moscow

Natalia V. Tarbaeva

Endocrinology Research Centre

Email: ntarbaeva@inbox.ru
ORCID iD: 0000-0001-7965-9454
SPIN-code: 5808-8065

MD, Cand. Sci. (Medicine)

Russian Federation, Moscow

Anna A. Roslyakova

Endocrinology Research Centre

Email: aroslyakova12@gmail.com
ORCID iD: 0000-0003-1857-5083
SPIN-code: 5984-4175

MD

Russian Federation, Moscow

Dmitry G. Beltsevich

Endocrinology Research Centre

Email: belts67@gmail.com
ORCID iD: 0000-0001-7098-4584
SPIN-code: 4475-6327

MD, Dr. Sci. (Medicine)

Russian Federation, Moscow

Liliya S. Urusova

Endocrinology Research Centre

Email: liselivanova89@yandex.ru
ORCID iD: 0000-0001-6891-0009
SPIN-code: 5151-3675

MD, Dr. Sci. (Medicine)

Russian Federation, Moscow

Natalia G. Mokrysheva

Endocrinology Research Centre

Email: mokrisheva.natalia@endocrincentr.ru
ORCID iD: 0000-0002-9717-9742
SPIN-code: 5624-3875

MD, Dr. Sci. (Medicine), Professor

Russian Federation, Moscow

Valentin E. Sinitsyn

Lomonosov Moscow State University; Research and Practical Clinical Center for Diagnostics and Telemedicine Technologies

Email: vsini@mail.ru
ORCID iD: 0000-0002-5649-2193
SPIN-code: 8449-6590

MD, Dr. Sci. (Medicine), Professor

Russian Federation, Moscow; Moscow

References

Supplementary files