Classification of adrenocortical carcinoma, pheochromocytoma, and adrenal adenomas using contrast-enhanced computed tomography with machine learning and texture features: a cross-sectional study

Full Text

BACKGROUND

Differential diagnosis of adrenal lesions remains one of the key challenges of modern oncologic radiology. Contrast-enhanced computed tomography (CT) is recognized as the primary method for adrenal tumor imaging due to high spatial and contrast resolution [1]. Both non-contrast CT and contrast-enhanced CT are used to differentiate between benign and malignant adrenal lesions [2]. The European Society of Endocrinology guidelines report six studies on the diagnostic accuracy of non-contrast CT [3–8]. It is noted that an X-ray density value > 10 HU is characterized by 100% sensitivity in the diagnosis of malignant lesions and, at the same time, a relatively low average specificity (58%). The diagnostic accuracy of CT with contrast washout assessment in the differentiation of malignant and benign adrenal lesions was studied by W. Schloetelburg et al. [6]. Thus, the authors determined that with a relative washout ratio of 58%, it is possible to achieve 100% sensitivity, whereas specificity was 15%. Thus, it can be concluded that the current CT protocol failes to reliably distinguish for a convincing differentiation between benign and malignant adrenal lesions [adrenocortical carcinoma (ACC), pheochromocytoma] due to the substantial overlap in their CT imaging characteristics.

CT limitations create serious clinical risks. In the case of ACC, for instance, surgical treatment is required [9], whereas adenomas are amenable for active surveillance [10–13]. Conversely, pheochromocytomas may exhibit CT semiotics similar to adenomas and ACC [14]. In such conditions, radiomics, a method of quantitative analysis of texture and morphological features on medical images, becomes a promising tool for overcoming the limitations of standard CT, providing a quantitative assessment of adrenal tumor heterogeneity [15], which is also noted in the recommendations of the European Society of Endocrinology [2]. A systematic review of 28 studies demonstrates the potential capabilities of CT radiomics in the differential diagnosis of adrenal tumors, providing high diagnostic accuracy [(Area Under the Curve, AUC) — 0.88] [16]. However, the authors of the review note the high heterogeneity of the included studies, raising doubts about the reliability of these results. Moreover, subgroup analysis did not reveal advantages for studies that used machine learning compared with those that did not. At the same time, because the number of such works was small (n = 3), the authors of the meta-analysis concluded that further research is needed for an objective assessment of the effectiveness of machine learning methods in adrenal radiomics. In another systematic review by Ferro et al. [17] also noted that the studies demonstrate the high diagnostic potential of CT radiomics: an AUC value of 0.88 for differentiating benign and malignant adrenal lesions, as well as functional and non-functional adrenal tumors. The authors specifically emphasize the greater effectiveness of CT compared with magnetic resonance imaging. Despite the encouraging results, challenges remain such as small sample sizes and a lack of standardization. In a study involving patients with malignant lung lesions, a machine learning model (based on the support vector machine method) using radiomics features successfully differentiated adrenal metastases and benign lesions (AUC = 0.938), surpassing traditional clinical-radiological approaches (assessment of hormonal activity, qualitative assessment of non-contrast CT results, and contrast washout rates) [18]. In another study dedicated to the differentiation of functioning and non-functioning adrenal adenomas, using logistic regression based on texture features of non-contrast, arterial, and venous phase CT images, a diagnostic accuracy of up to 83% was achieved. The obtained results indicate the potential of the method to minimize the use of invasive tests for diagnostic purposes [19].

Despite progress, practically all studies in this area are focused on binary classification (benign and malignant) or narrow subgroups (for example, differentiation of adenomas into types). At the same time, complex models for the simultaneous diagnosis of three key classes of adrenal lesions (adrenal adenomas, ACC, and pheochromocytoma) have not been developed. A study conducted by Tucci et al. [20] is the only one in which the classification of adrenal tumors into more than 2 classes was considered. However, its results are preliminary and have been published as conference proceedings. Therefore, the model architecture, its parameters, and the characteristics of the study cohort remain unclear. At the same time, the multiclass classification approach seems fundamentally substantial to us, as pheochromocytomas, despite their potentially malignant nature, require a diagnostic and therapeutic strategy that differs from that for ACC and adenomas [2].

AIM

To develop a machine learning model for multiclass classification of adrenal lesions (adenomas, ACC, and pheochromocytomas) using contrast-enhanced computed tomography data with texture features.

METHODS

Study Design

This was a single-center, cross-sectional study with retrospective computed tomography data acquisition and prospective re-analysis of imaging results.

Data Source

The study included data from patients registered in the medical information system of the I. I. Dedov National Medical Research Center of Endocrinology (Moscow) in the period from January 2018 to December 2024.

Selection Criteria

Inclusion criteria:

• fact of adrenal lesion removal;
• histologically confirmed diagnosis of adenoma, ACC or pheochromocytoma;
• availability of four-phase contrast-enhanced CT performed at the I. I. Dedov National Medical Research Center of Endocrinology.

Non-inclusion criteria:

• exclusion criteria: presence of artifacts in the adrenal region on CT images (motion artifacts, ring artifacts);
• equivocal results of pathomorphological examination.

Exclusion criteria: none planned.

Study Outcome

The study outcome (predicted event) was considered to be the presence of an adrenal lesion (adenoma, ACC, or pheochromocytoma) verified by the results of histological examination of postoperative material using the standard procedure in the Department of Fundamental Pathomorphology of the I. I. Dedov National Medical Research Center of Endocrinology.

Measurement of Study Outcomes

Adrenal lesion samples were obtained during elective adrenalectomies. Surgical material was labeled and placed in 10% buffered formalin with a fixative-to-tissue volume ratio of approximately 10:1. Fixation time was 24–48 hours. Subsequently, the samples underwent standard histological processing using a Leica® ASP6025 S tissue processor (Leica Biosystems, Germany), followed by paraffin embedding. Equatorial sections were prepared from the paraffin blocks using a Leica® RM 2125 RTS microtome (Leica Biosystems, Germany). Sections were deparaffinized and stained with hematoxylin and eosin using a Leica® ST5010 AXL stainer (Leica Biosystems, Germany). Morphological study of the slides was performed by light microscopy using a Leica® DM2500 microscope (Leica Microsystems, Germany). Aperio® ImageScope software (Leica Microsystems, Germany) was used for quantitative morphometry. Morphological assessment of adrenocortical tumors was performed in accordance with generally accepted histopathological criteria:

• size and weight of adrenal tumor;
• degree of nuclear pleomorphism;
• mitotic activity (mitotic count in 10 fields of view at × 400 magnification);
• characteristics of tumor cell cytoplasm (clear in 0–25 or 26%–100% of cases);
• architectural growth pattern (diffuse or non-diffuse);
• presence of atypical mitosis, necrotic foci, as well as signs of invasion — capsular, venous, sinusoidal, and invasion into adjacent structures.

The malignant potential of adrenocortical tumors was determined according to the Weiss Functional Impairment Rating Scale [21]. In the case of diagnosing oncocytic tumors with characteristic granular, intensely eosinophilic cytoplasm, pronounced nuclear polymorphism, and a diffuse growth pattern, the modified Lin–Weiss–Bisceglia scale was applied [22]. The criterion for adrenocortical carcinoma is ≥ 4 features on the Weiss scale.

Mitotic activity was determined by counting the number of mitotic figures in 50 high-power fields (×400) using a Leica® DM2500 microscope (Leica Microsystems, Germany). Counting was carried out in areas with the highest density of mitosis. Whenever possible, high-power fields were selected from different slides. Diffuse tumor architecture was defined when more than 33% of the cross-sectional area was composed of growth areas without a clear organoid structure.

To assess vascular invasion, a distinction was made between veins — endothelium with a pronounced muscular layer and sinusoids, lined with endothelium but lacking a significant muscular coat. 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.

The malignant potential of pheochromocytoma was assessed using the PASS (Pheochromocytoma of the Adrenal Gland Scaled Score) scale, which includes 12 main morphological features. With a total score of ≥ 4, the malignant potential of the adrenal tumor was assessed as high [23].

Differential diagnosis of adrenal tumors in all cases was performed in the Department of Fundamental Pathomorphology based on the results of immunohistochemical study of the same samples. The main immunohistochemical panel included markers of cortical tumor histogenesis — steroidogenic factor 1, inhibin A, and melan-A, as well as markers of neuroendocrine differentiation — chromogranin A, synaptophysin (Leica kits, Germany, were used to determine all markers). In all cases of ACC and adrenocortical adenomas, positive staining was detected using antibodies to steroidogenic factor 1, which has the highest sensitivity (98%) and specificity (100%), and is also a criterion for adrenocortical tumors [24].

Throughout the study (2018–2024), there were no changes in the methodology of histological and immunohistochemical studies.

Contrast-Enhanced Computed Tomography Data

The results of preoperative contrast-enhanced CT performed both in the I. I. Dedov National Medical Research Center of Endocrinology and in other medical institutions according to the peritoneal cavity study protocol were analyzed. The latter, in CD format, were analyzed by radiologists of the Department of Computed and Magnetic Resonance Tomography of the I. I. Dedov National Medical Research Center of Endocrinology in the period from November 2024 to July 2025. All contrast-enhanced CT scans were performed in the period from 1 day to 1 month before surgical intervention.

Computed Tomography Acquisition

CT scans in the I. I. Dedov National Medical Research Center of Endocrinology between 2018 and 2024 were performed concurrently on Optima® CT660 and Revolution® CT scanners (GE Healthcare, USA) with various scanning parameters depending on the phase (non-contrast, arterial, venous, delayed). Main parameters for the Optima® CT660 scanner (GE Healthcare, USA): tube voltage — 100–120 kV; slice interval — 0.625 mm; slice thickness — 1.25 mm. For the Revolution® CT scanner (GE Healthcare, USA): tube voltage — 100–140 kV; slice interval — 0.625–1.25 mm; slice thickness — 0.625–1.25 mm. 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.

Image Post-Processing

Segmentation of adrenal lesion regions on CT images was performed using open-source software 3D Slicer® 5.6.2 (Slicer Community, USA) for each scanning phase. After segmentation, texture features were calculated using the Python 3.9.21 programming language and the PyRadiomics® 3.1.0 module (Computational Imaging & Bioinformatics Lab, USA). A calculation of 106 features (for each computed tomography phase) from several groups was performed:

• first-order statistical features — based on the histogram of the intensity distribution of texture elements in the region of interest (ROI);
• second-order statistical features — based on the Gray Level Run Length Matrix (GRLRM) — 16 features; based on the Neighbouring Gray Tone Difference Matrix (NGTDM) — 5 features; based on the Gray Level Dependence Matrix (GLDM) — 14 features; based on the Gray Level Size Zone Matrix (GLSZM) — 16 features.

The binWidth parameter was set to 5. Features were calculated for the original CT images; after applying a Laplacian of Gaussian filter that standard deviation was set to 1, 3, and 5 mm to highlight boundaries in the region of interest; after wavelet transform for decomposition of CT images into components of various spatial frequencies (the "coif1" wavelet was used). All other PyRadiomics module parameters regulating image post-processing and feature calculation were left at default. All parameters and descriptions of texture features have been published by the developers of the PyRadiomics module [25].

Data Preparation

Assessment of Feature Reproducibility and Variability, Data Standardization

To assess texture features reproducibility, radiologists from two groups: one with work experience of ≥5 years (n = 3) and the other with < 5 years of experience (n = 3) independently segmented a subset of adrenal lesions (45 cases). Segmentation of adrenal lesions was performed in all phases of contrast-enhanced CT within the study. Next, we calculated texture features for the segmented regions of interest, and compared their values based on rater category. The Intraclass Correlation Coefficient (ICC) was used as a metric for feature reproducibility. An ICC value > 0.9 was accepted as the reproducibility criterion.

To assess the variability of texture features, a radiologist with more than 5 years of work experience in the radiologic diagnosis of adrenal lesions delineated a region of interest within the region of the aorta in the non-contrast CT phase for a subsample of 30 patients. After that, texture features were calculated for it. To assess variability, the coefficient of variation (CV) was used. CV values < 0.15 were considered a criterion for low variability. The radiologist did not have access to the patients' clinical data, to the interpretation results of contrast-enhanced CT images, or to the pathomorphological findings.

Data standardization method was performed due to the variability of CT scanning parameters in the sample and their impact on texture features [26], using the region inside the aorta as a reference structure. Assuming that feature variability is related to the variance of texture element intensities in the region of interest, scaling was performed according to the following formula:

$I_{standardized}(x, y, z)={\mu }_{image}+\left(I\right(x, y, z)-{\mu }_{image})\times \frac{{\sigma }_{ref}}{{\sigma }_{aorta}}$, (1)

where I (x, y, z) is the initial X-ray density value in a voxel with coordinates (x, y, z); μ_image is the mean X-ray density in the region of interest; σ_aorta is the standard deviation of X-ray density in the region of the aorta on the image; σ_ref is the reference standard deviation of density in the region of the aorta, defined as the mean σ_aorta value across the entire sample (20 HU).

Feature Selection for the Machine Learning Model

Before training the models, a feature selection stage was performed, excluding highly correlated features (> 0.9 according to Spearman correlation coefficient). In pairs of such features, the one characterized by a lower MaxSD metric value, determined according to the formula, was excluded:

$MaxSD=\frac{\max \left(\left|{\mu }_i-{\mu }_j\right|\right)}{{\sigma }_{pooled}}$, (2)

where μ_i, μ_j are the mean values of the feature for different diagnoses (ACC, adenomas, pheochromocytomas); σ_pooled is the standard deviation of the feature in the entire sample, including all diagnoses.

Machine Learning Model

The machine learning model was developed based on gradient boosting models (XGBoost, LightGBM, CatBoost models; lightgbm, xgboost, and catboost modules in the Python language, respectively). Texture features extracted from contrast-enhanced CT images were used as outcome predictors. The gradient boosting model predicts the probability score of the tumor belonging to the category of adrenal lesions (ACC, adenoma, pheochromocytoma). To convert the continuous probability into a categorical model output, classification was established as selecting the class with the maximum probability value.

Model Hyperparameters

To optimize model complexity and minimize the risk of overfitting, the Early Stopping method was used during training (the number of iterations without a decrease in the loss function value on the validation sample, at which the training cycle stops, was set to 20). The following hyperparameter values were set:

• max_depth = 3 (maximum depth of the base decision tree algorithm);
• learning_rate = 0.1 (responsible for updating model parameters during training).

The logistic loss function for multiclass classification (categorical cross-entropy) was used as the loss function.

Class Balancing

To minimize class imbalance in terms of the number of observations, class weighting was performed during training (assigning each class a weight inversely proportional to the number of observations in that class).

Model Evaluation

Diagnostic performance was assessed using Recall (sensitivity), Precision, F1-score, macro-averaged F1-score, specificity, Balanced Accuracy, and AUC for each diagnostic category. Values of all metrics above 0.80 were considered a criterion for high diagnostic accuracy of the model. The final values were averaged using stratified five-fold cross-validation to obtain an estimate of the model's generalization ability. The standard deviation calculated by the cross-validation was used to assess model variability. It was assumed that Balanced Accuracy and macro-averaged F1-score are reliable indicators of diagnostic performance under conditions of class imbalance [27, 28].

Evaluation of Model Feature Importance

To evaluate the informativeness of texture features (i.e., their contribution to the predictive ability of the models), feature importance was assessed using the built-in methods of each algorithm. As various gradient boosting frameworks implement distinct importance metrics, feature importance for the LightGBM model was assessed using the Gain metric at data splits. The XGBoost model also used the Gain metric, reflecting the average reduction in the loss function achieved when using a feature. For the CatBoost classifier, the PredictionValuesChange was applied.

Sample Splitting

The sample was randomly split into training and test sets in an 80:20 ratio with stratification using the train_test_split function of the sklearn module in the Python programming language. The final model was trained on the full training set before metrics were computed on the test sample.

Ethics Approval

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

Statistical Analysis

Sample size justification. The samples size was not calculated previously.

Statistical methods. Python, version 3.9.21, was used for data analysis. Data for continuous variables [age at the time of contrast-enhanced CT, maximum linear size of the adrenal lesion based on CT data (mm), CT attenuation across phases in Hounsfield units (HU)] are presented as Me [Q1; Q3], where Me is the median, and Q1 and Q3 are the 1st and 3rd quartiles, respectively. To compare the values of categorical features across different diagnoses, the χ² criterion was used, and for continuous variables, the Kruskal–Wallis test (with Dunn's test as a post-hoc test) with the Bonferroni correction for multiple hypothesis testing.

RESULTS

Sampling

Surgical treatment for adrenal lesions in the Department of Surgery of the I. I. Dedov National Medical Research Center of Endocrinology in period 2018 – 2024, 847 patients were observed, all of them with histological examination results. Preoperative contrast-enhanced CT was available in the medical information system in 466 cases; in 381 cases, results with all four phases were missing. Data from 30 patients with indeterminate histological results and 11 patients with motion and ring artifacts were excluded from the study. As a result, a final sample included data from 425 patients with histologically verified adrenal tumors: 42 cases of ACC, 204 pheochromocytomas, and 179 adrenal adenomas. A complete set of planned data was obtained for all patients included in the study.

Characteristics of Study Groups

Patient groups with adrenal lesions were comparable by sex and age (see Table 1). Group differences were observed in maximum linear size (adrenal tumors in ACC were considerably larger compared with pheochromocytomas and adenomas according to pairwise comparison results), density in the native CT phase (adenomas were characterized by lower density than ACC and pheochromocytomas), and density in the arterial, venous, and delayed phases (pheochromocytomas demonstrate the highest density values in dynamic contrast phases, whereas there are no significant differences between ACC and adenomas).

 

Table 1. Characteristics of patients with adrenal lesions

Parameter	Adrenocortical carcinoma, n = 42	Pheochromocytoma, n = 204	Adenoma, n = 179	p	p, post-hoc
Male sex, n (%)	16 (37)	80 (38)	65 (35)	0.832	p12 = 1.000 p13 = 1.000 p23 = 1.000
Age, years	48.0 [33.4; 57.7]	49.8 [39.0; 59.3]	49.0 [41.0; 59.2]	0.380	p12 = 0.189 p13 = 0.176 p23 = 0.936
Maximum linear size, mm	78.0 [58.0; 99.5]	47.5 [34.0; 62.0]	30.0 [21.0; 39.0]	< 0.001	p12 < 0.001 p13 < 0.001 p23 < 0.001
Attenuation in non-contrast phase, HU	44.0 [41.5; 47.5]	48.0 [42.5; 53.0]	30.0 [17.0; 40.0]	< 0.001	p12 = 0.048 p13 < 0.001 p23 < 0.001
Attenuation in arterial phase, HU	83.5 [62.0; 127.3]	126.0 [110.0; 166.0]	79.0 [57.0; 102.0]	< 0.001	p12 < 0.001 p13 = 0.070 p23 < 0.001
Attenuation in venous phase, HU	101.0 [81.0; 113.5]	114.0 [97.0; 137.0]	100.0 [80.8; 125.3]	< 0.001	p12 = 0.003 p13 = 0.720 p23 < 0.001
Attenuation in delayed phase, HU	64.0 [59.8; 72.0]	71.5 [63.0; 83.3]	58.0 [43.0; 75.0]	< 0.001	p12 = 0.012 p13 = 0.144 p23 < 0.001

Note. The results are presented as Me [Q1; Q3], where Me is the median and Q1 and Q3 are the first and third quartiles, respectively. To compare the values of categorical features across different diagnoses, the χ² test was used, and for continuous variables, the Kruskal–Wallis test (with Dunn's test as a post-hoc test) was used.

 

Primary Results

Assessment of Reproducibility and Variability of Texture Features

After segmentation of adrenal lesions by two categories of radiologists and calculation of texture features, the ICC for various CT phases was determined. The proportion of features with an ICC value > 0.9 varied in 70% to 82% (see Table 2).

 

Table 2. Reproducibility of texture feature assessments

Computed tomography phase	ICC > 0,9, n (%)
Non-contrast	892 (78.9)
Arterial	907 (80.3)
Venous	794 (70.3)
Delayed	930 (82.3)

Note. The total number of analyzed texture features is 1130. ICC (Intraclass Correlation Coefficient) — intraclass correlation coefficient.

 

After segmentation of the region inside the aorta and calculation of CV, the proportion of low-variability features with CV < 0.15 was 12.9% (146/1130). After applying standardization using the region inside the aorta as a reference structure, a CV value < 0.15 was determined for 313/1130 (27.7%) features. For further analysis, we used only 245 low-variability and reproducible features after the selection procedure.

Quality of Models on Test and Validation Data

The test sample included 85 cases of adrenal lesions, and the training set included 340. Training of XGBoost, CatBoost, and LightGBM models was performed. The final performance metrics based on the test sample and the results of the cross-validation are presented in Table 3. According to the Balanced Accuracy and macro-averaged F1-score values, within the variability assessment (standard deviation), the models are characterized by similar values with overlapping metric values. At the same time, the Recall, Precision, and F1-score values for ACC are substantially lower compared with the corresponding metrics for adenomas and pheochromocytomas. However, on the test sample, the metric values for all three models are identical (aside from a small difference in AUC values).

 

Table 3. Model performance metrics based on cross-validation data and evaluation on the test sample

Metrics	Gradient boosting model
	LightGBM	XGBoost	Catboost
Cross-validation results
Adenomas
Recall	0.866 ± 0.041	0.880 ± 0.029	0.859 ± 0.067
Precision	0.949 ± 0.044	0.958 ± 0.051	0.948 ± 0.045
F1-score	0.904 ± 0.026	0.916 ± 0.029	0.900 ± 0.042
Specificity	0.964 ± 0.031	0.969 ± 0.038	0.964 ± 0.031
AUC	0.971 ± 0.015	0.967 ± 0.021	0.959 ± 0.021
Pheochromocytomas
Recall	0.884 ± 0.036	0.909 ± 0.066	0.817 ± 0.060
Precision	0.810 ± 0.018	0.810 ± 0.012	0.831 ± 0.043
F1-score	0.845 ± 0.022	0.855 ± 0.033	0.822 ± 0.033
Specificity	0.806 ± 0.021	0.800 ± 0.018	0.840 ± 0.056
AUC	0.924 ± 0.022	0.920 ± 0.027	0.917 ± 0.023
Adrenocortical carcinoma
Recall	0.452 ± 0.169	0.362 ± 0.120	0.643 ± 0.136
Precision	0.503 ± 0.163	0.506 ± 0.119	0.450 ± 0.077
F1-score	0.464 ± 0.142	0.409 ± 0.099	0.521 ± 0.075
Specificity	0.951 ± 0.021	0.961 ± 0.016	0.912 ± 0.030
AUC	0.878 ± 0.054	0.871 ± 0.052	0.879 ± 0.063
Aggregated performance metrics
Balanced Accuracy	0.734 ± 0.046	0.717 ± 0.028	0.773 ± 0.040
Macro-averaged F1-score	0.738 ± 0.046	0.727 ± 0.033	0.747 ± 0.038
Test dataset
Adenomas
Recall	0.861	0.861	0.861
Precision	0.912	0.912	0.912
F1-score	0.886	0.886	0.886
Specificity	0.939	0.939	0.939
AUC	0.976	0.982	0.974
Pheochromocytomas
Recall	0.902	0.902	0.902
Precision	0.787	0.787	0.787
F1-score	0.841	0.841	0.841
Specificity	0.773	0.773	0.773
AUC	0.946	0.950	0.947
Adrenocortical carcinoma
Recall	0.375	0.375	0.375
Precision	0.750	0.750	0.750
F1-score	0.500	0.500	0.500
Specificity	0.987	0.987	0.987
AUC	0.940	0.953	0.930
Aggregated performance metrics
Balanced Accuracy	0.713	0.713	0.713
Macro-averaged F1-score	0.742	0.742	0.742

Note. AUC (Area Under the Curve) — area under the curve. Values of metrics obtained as a result of the cross-validation are presented as M ± SD, where MD is the mean value and SD is the standard deviation.

 

The most informative texture features for classification [the first five features by their importance (Feature Importance)] for each model, along with their brief interpretation, are presented in Table 4. The most important texture feature for solving the classification task is wavelet-LLL_firstorder_Mean_NAT, which reflects the mean X-ray density in the adrenal tumor region on the CT image in the native phase. The distribution of the mean values of these features for different categories (diagnoses) is presented in Fig. 1. The distribution of observations according to the predicted class showed that the models are characterized by a low ability to differentiate between ACC and pheochromocytomas (Fig. 2).

 

Table 4. Significant texture features and their interpretation

Feature	Feature importance, arbitrary units	Brief feature interpretation
LightGBM
Wavelet-LLL_firstorder_Mean_NAT	71	Reflects the average X-ray density in the region of interest after removing high-frequency noise and preserving global (low-frequency) structural changes in the native phase computed tomography image.
Log-sigma-3-0-mm-3D_glcm_MCC_ART	37	MCC (Matthews correlation coefficient) describes the linear dependence of image texture elements after applying the Laplacian of Gaussian filter at a detail level of ~3 mm in the arterial phase of computed tomography; a high MCC on the LoG image describes complex, irregular, invasive tumor boundaries, which visually often appear less clear (blurred, jagged) due to their complex structure. Low MCC is associated with smooth, well-defined boundaries.
Original_shape_Maximum2DDiameterRow_VEN	29	Maximum tumor diameter in the axial plane in the venous phase of computed tomography.
Wavelet-LLL_firstorder_90Percentile_DEL	23	90th percentile of X-ray density values in the region of interest after removing high-frequency noise and preserving global (low-frequency) structural changes in the delayed phase.
Wavelet-LLL_firstorder_Mean_ART	21	Reflects the mean X-ray density in the region of interest after removing high-frequency noise and preserving global (low-frequency) structural changes on the arterial phase computed tomography image.
XGBoost
Original_glrlm_ShortRunEmphasis_VEN	14	Reflects the predominance of short runs of equal intensities in the region of interest; a higher feature value indicates a fine-grained and finely fragmented texture.
Wavelet-LLL_firstorder_Mean_NAT	13	See interpretation above (in the LightGBM model block).
Log-sigma-1-0-mm-3D_firstorder_10Percentile_VEN	11	10th percentile for the region of interest after applying the Laplacian of Gaussian filter; sensitive to a more uniform distribution of boundaries with a predominance of weak density changes; a higher value indicates pronounced boundaries.
Original_shape_Maximum2DDiameterRow_VEN	11	See interpretation above (in the LightGBM model block).
Wavelet-LLL_firstorder_Mean_ART	9	See interpretation above (in the LightGBM model block).
CatBoost
Wavelet-LLL_firstorder_Mean_NAT	14	See interpretation above (in the LightGBM model block).
Wavelet-LLL_firstorder_Mean_DEL	14	Reflects the average X-ray density in the region of interest after removing high-frequency noise and preserving global (low-frequency) structural changes in the delayed phase computed tomography image.
Original_shape_Maximum2DDiameterRow_VEN	8	See interpretation above (in the LightGBM model block).
Wavelet-LLL_firstorder_Mean_ART	4	See interpretation above (in the LightGBM model block).
Original_firstorder_90Percentile_NAT	3	90th percentile of X-ray density values in the native phase.

 

Fig. 1. Heat map of texture feature values for different adrenal lesions. The standardized mean value (equal to the ratio of the difference between the mean value of a specific diagnosis and the mean value of the entire sample to the standard deviation of the entire sample) and the original mean value (in parentheses) of the corresponding texture feature indicator are shown. ACC — adrenocortical carcinoma.

 

Fig. 2. Accuracy of observed classification of test sample (for all three models considered): 0 — adenomas; 1 — pheochromocytomas; 2 — adrenocortical carcinoma.

 

DISCUSSION

Interpretation of Study Results

The developed machine learning models based on texture features of contrast-enhanced CT images are characterized by similar quality metrics. They demonstrated high classification performance for pheochromocytomas and adrenal adenomas, but not for ACC. The impact of differences in CT scanning parameters was minimized by data standardization using the region inside the aorta as a reference structure.

Discussion of Study Results

The results obtained demonstrate that the proposed gradient boosting models based on CT image texture features provide high diagnostic accuracy for predicting pheochromocytomas and adrenal adenomas, consistent with previously published studies focused on the binary classification of these pathologies. For example, a meta-analysis by Zhang et al. [16] showed a mean AUC = 0.880 among investigations for the differentiation of benign and malignant adrenal tumors using texture features, whereas in the work by Cao et al. [18] achieved an AUC of 0.938 in differentiating between metastases and benign lesions.

However, unlike these works, our model is focused on the multiclass classification of three groups of tumors (ACC, pheochromocytomas, and adenomas), which is a more complex task. According to the results of our study, the models were unable to correctly classify ACC. This may be accounted for by specific factors, namely:

• ACC heterogeneity — it is known to demonstrate considerable structural and functional heterogeneity [29], which complicates its identification based on texture features;
• limited sample size — in our study, ACC is represented by 42 cases, which is substantially smaller than the number of pheochromocytomas and adrenal adenomas;
• class imbalance, affecting classification accuracy;
• overlap of texture characteristics — some texture features of ACC are similar to pheochromocytomas, which can lead to misclassification.

Comparison with a large multicenter study [20], which used a neural network model with texture features, reveals key differences. In the aforementioned work, high accuracy was achieved for the classification of malignant lesions (combined into one group: ACC, pheochromocytomas, metastases), with an AUC = 0.974, F1-score = 0.801. Moreover, the accuracy of ACC classification in the study (AUC = 0.973) substantially exceeds our results. This discrepancy may be due to the fact that the use of a neural network model in the study by Tucci et al. [20] could have provided better detection of complex ACC texture patterns. In the present study, we used gradient boosting models, which are potentially inferior to neural network algorithms when processing big data. Neural networks are capable of approximating any continuous function, identifying complex surfaces in the feature space, whereas gradient boosting, relying on piecewise constant trees and step-by-step training, has a limited ability to approximate complex dependencies [30, 31]. However, in contrast to neural network models, gradient boosting models require less training data, offer direct quantitative assessment of feature importance, and demonstrate resistance to overfitting [32]. Furthermore, Tucci et al [20] observed that adding data on tumor hormone secretion (adding a categorical variable to the model with possible values: non-secreting; mild autonomous cortisol secretion; Cushing syndrome; catecholamine-producing tumor) substantially improved classification accuracy (area under the curve = 0.999), which confirms the importance of integrating clinical data with texture characteristics. In our study, we did not use such information, which could also have affected the classification accuracy.

Special attention should be paid to the texture features that demonstrated the highest informativeness for our task. The predominant feature was the wavelet-LLL_firstorder_Mean_NAT feature, representing the mean X-ray attenuation following the elimination of high-frequency spatial component and preserving global (low-frequency) structural changes. ACC and pheochromocytomas are characterized by an on average high value of this feature compared with adrenal adenomas, which is associated with the known fact of lower X-ray density of adenomas in the native phase [33]. The same fact also explains the lower value of the original_firstorder_90Percentile_NAT feature for adrenal adenomas compared with ACC and pheochromocytomas. Also, ACC is characterized by on average the largest size of lesions (original_shape_Maximum2DDiameterSlice_VEN feature), which is consistent with the results of the study by Robertson-Tessi et al [34]. Furthermore, for ACC, the values of the log-sigma-3-0-mm-3D_glcm_MCC_ART and log-sigma-1-0-mm-3D_firstorder_10Percentile_VEN features are on average higher, which indicates the heterogeneity of its structure compared with pheochromocytomas and adrenal adenomas. Importantly, that the original_glrlm_ShortRunEmphasis_VEN feature also emerged as significant in our study; ACC is characterized by an on average lower value of it than adenomas and pheochromocytomas, which indicates more pronounced large-scale heterogeneity (likely areas of necrosis, calcifications) of ACC, as previously described [34]. At the same time, the values of the features wavelet-LLL_firstorder_90Percentile_DEL, wavelet-LLL_firstorder_Mean_ART, and wavelet-LLL_firstorder_Mean_DEL were lower in ACC than in pheochromocytomas, whereas adrenal adenomas had the lowest values among all studied classes. This primarily indicates a higher X-ray density in the arterial and delayed phases for pheochromocytomas compared with ACC, as well as faster contrast washout by adenomas [35]. A higher X-ray density in the arterial phase for pheochromocytomas compared with ACC was also observed in the study by Phadte et al. [36], which is consistent with our results. Adrenal adenomas in our sample generally demonstrate a more homogeneous texture. In the study by Altay et al. [37], adenomas are characterized by a shorter run length of bright voxels, which indicates their more homogeneous structure compared with malignant adrenal tumors. These results allow for the conclusion that the key textural characteristics for differentiating adrenal tumor types are features reflecting overall brightness, homogeneity, and local intensity variations.

This study identified a systematic classification error in which models more frequently misclassified ACC cases as pheochromocytomas, but not vice versa, which is likely due to several interrelated factors. First, class imbalance in the training sample is a key limiting factor. ACC is classified as an orphan disease; therefore, the number of ACC cases in the sample is substantially smaller than that of adenomas and pheochromocytomas. In such a situation, reliable identification of features specific to ACC becomes problematic, whereas patterns characteristic of the more numerous group of pheochromocytomas are better identified by the models. As a result, models "lean" toward a more frequent, "safe" prediction from the perspective of their internal optimization in favor of pheochromocytomas when performing an analysis of questionable cases. Second, analysis of the most relative features indicates that parameters such as mean X-ray density (wavelet-LLL_firstorder_Mean) and maximum tumor diameter are most significant for classification. ACC and pheochromocytomas are both known to commonly present as large lesions, with heterogeneous texture, with areas of necrosis and hemorrhage, which can lead to substantial similarity in their semiotics according to CT data [14]. Thus, these two classes are closely positioned within the multidimensional space of texture features. The fact that the inverse error occurs less frequently can be explained by the fact that pheochromocytomas, being a more represented class, form a more compact and well-defined cluster in the feature space. The model more reliably identifies their "typical" cases. In contrast, ACC, due to its rarity and greater morphological variability, represents a more "diffuse" class.

In addition, the study conducted a comparative evaluation of the effectiveness of three gradient boosting models — LightGBM, XGBoost, and CatBoost — for the classification of adrenal lesions based on CT image texture features. The key point is that all three models demonstrated identical results on the test sample. This confirms that under the conditions of this dataset, the models has comparable generalization capability and demonstrate stable work. Additionally, the cross-validation analysis, which enables evaluation of model variability, revealed that the performance metrics for all models overlapped within the standard deviation. This does not allow for an unequivocal conclusion regarding the statistically significant superiority of any gradient boosting model. The results of the published studies comparing these three models are also ambiguous and show that there is no universally superior implementation of gradient boosting: LightGBM provides an optimal balance between accuracy and computational efficiency after tuning; CatBoost outperforms analogs when working with categorical features and is characterized by stability; XGBoost remains a reliable and flexibly tunable base model. The discrepancies among the results are likely attributable to characteristics of the data used, evaluation methods, and optimization strategies [38].

An important direction for increasing diagnostic accuracy could be a narrower formulation of the classification task, in particular, a focus on the differentiation of pheochromocytomas and ACC, because of substantial overlap in their radiologic characteristics. Preliminary filtering and the analysis of only the solid components of the adrenal tumor, with the exclusion of areas of necrosis and cystic inclusions, may improve the informativeness of texture features by reducing the impact of such heterogeneities on radiomic metrics. Such an approach could potentially enhance the ability of models to recognize stable texture patterns characteristic of pheochromocytomas and adenomas, which, in turn, could improve their sensitivity and specificity. At the same time, this approach requires additional research.

Study Limitations

One of the limitations of the present study is the relatively small number of ACC cases (n = 42), which may reduce the stability and reproducibility of the constructed models, increasing the risk of their overfitting during the analysis of this group. Although the presented distribution of classes (diagnoses) reflects the actual prevalence of various adrenal tumors in our sample (operated patients with a verified morphological diagnosis), it introducers substantial data imbalance, a critical consideration in the construction and evaluation of machine learning models. In particular, when training classification models, we applied class weighting, in which each class was assigned a weight inversely proportional to its frequency, in order to minimize model bias toward the predominant classes and increase sensitivity to less common adrenal tumors, such as ACC.

The use of only texture features without considering clinical data (for example, the level of hormonal activity) could limit the capabilities of the models in complex cases [20]. Also, the study results require validation in independent patient cohorts to confirm their reproducibility and clinical applicability.

Notebly, that the reference test in the present work was the histopathological examination used as the "gold standard" in diagnostic studies [2]. This method has high sensitivity and specificity in establishing the final diagnosis [39]. Nevertheless, histopathological diagnostics may also be subject to certain limitations, including subjective interpretation of histological criteria (e.g. the Weiss scale) and difficulties in diagnosing rare tumors [40]. This factor could have affected the evaluation of the diagnostic accuracy of machine learning models and its generalizability. The impact is reflected in inter-observer variability among pathologists, which creates "noise" in the reference data on which the model is trained and validated, potentially compromising the accuracy of its performance.

Practical Significance

The results obtained in this study highlight the potential of the developed machine learning models based on texture features of contrast-enhanced CT data as an auxiliary tool in clinical practice for the preliminary differential diagnosis of adrenal lesions. High values of performance metrics for the classification of adenomas and pheochromocytomas indicate a good discriminative ability of the approach in distinguishing benign lesions from potentially malignant lesions, which is particularly important at the stages of primary Visualization and decision-making regarding the necessity of surgical intervention. Especially valuable is the model's potential to exclude adenomas, which often do not require invasive treatment, from the group of those suspicious for malignant lesions. However, the models demonstrated relatively poor performance in differentiating ACC from pheochromocytomas, limiting their utility for accurate nosological classification of two types of adrenal tumors. Nevertheless, even with existing limitations, the proposed models can be used as part of a comprehensive diagnostic algorithm alongside clinical, laboratory, and hormonal data to increase the validity of clinical decisions.

CONCLUSION

The use of machine learning and radiomics methods for multiclass classification of adrenal lesions based on contrast-enhanced CT data provides high accuracy in differentiating pheochromocytomas and adenomas, whereas the classification of ACC is complicated by its heterogeneity and limited sample size. The use of data standardization in the study reduced the impact of CT scanning parameters on the variability of texture features. The results obtained, comparable with previously published data, confirm the potential of radiomics in improving the differential diagnosis of adrenal tumors and indicate the need for further studies on larger samples.

ADDITIONAL INFORMATION

Author contributions: A.V. Manaev: conceptualization, data curation, formal analysis, visualization, methodology, writing — original draft, writing — review & editing; N.V. Tarbaeva: conceptualization, project administration, writing — original draft; S.A. Buryakina, L.D. Kovalevich, A.V. Khairieva, L.S. Urusova, N.V. Pachuashvili: data curation, investigation, resources, writing — review & editing; G.A. Mel'nichenko, 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 I. I. Dedov National Medical Research Center for Endocrinology (Minutes No. 20, dated November 13, 2024). All participants provided written informed consent for the use of their clinical assessment and treatment data for research purposes.

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: Original data from 24 patients with adrenocortical carcinoma were used in this study; these data had previously been used in an earlier published work (doi: 10.17816/DD643532) and are reproduced with permission of the copyright holder.

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 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
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

Svetlana A. Buryakina

Endocrinology Research Centre

Email: sburyakina@yandex.ru
ORCID iD: 0000-0001-9065-7791
SPIN-code: 5675-0651

MD, Cand. Sci. (Medicine)

Russian Federation, Moscow

Liliya D. Kovalevich

Endocrinology Research Centre

Email: liliyakovalevich@gmail.com
ORCID iD: 0000-0001-8958-8223
SPIN-code: 1642-5694
Russian Federation, Moscow

Angelina V. Khairieva

Endocrinology Research Centre

Email: komarito@mail.ru
ORCID iD: 0000-0002-6758-5918
SPIN-code: 4516-8297
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

Nano V. Pachuashvili

Endocrinology Research Centre

Email: npachuashvili@bk.ru
ORCID iD: 0000-0002-8136-0117
SPIN-code: 3477-8994

MD, Cand. Sci. (Medicine)

Russian Federation, Moscow

Galina A. Mel'nichenko

Endocrinology Research Centre

Email: Melnichenko.Galina@endocrincentr.ru
ORCID iD: 0000-0002-5634-7877
SPIN-code: 8615-0038

MD, Dr. Sci. (Medicine), Professor

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