RESEARCH
Differentiating between adrenocortical carcinoma and pheochromocytoma by a CT-based radiomics model: a multicenter retrospective study
Check for updates
Yinyao Chao41, Hongzhang Zhu5+, Wenyi Yang6+, Haohua Yao7+, Nan Ma1,2,3, Xianda Chen1,2,3, Jing Zhao2,3,8, Huali Ma2,3,8 Zhenhua Liu1,2,3, Hui Han1,2,3, Zhuowei Liu1,2,3,10, Kai Yao1,2,3, Yiyao Li9, Peng Wu9, Jingtong Zhang1,2,3*, Bin Li6* and Shengjie Guo1,2,3*
Abstract
Background Adrenocortical carcinoma (ACC), a rare and highly malignant adrenal gland tumor, exhibits computed tomography (CT) characteristics that resemble those of the less malignant pheochromocytoma (PHEO). While biochemical evaluation is widely accepted for differentiating between ACC and PHEO, non-functioning tumors remain a diagnostic challenge. The similarity in CT imaging and atypical hormone levels can lead to suboptimal accuracy in diagnosis, leading to inappropriate clinical interventions. This study aims to differentiate between large (≥4 cm) ACC and PHEO with radiomics features based on contrast-enhanced CT.
Methods In this retrospective study, 158 patients who received pathological diagnoses of ACC or PHEO between January 2011 and September 2023 were enrolled from three institutions. Radiomics features were extracted from different phases of contrast-enhanced CT and then selected by a two-step procedure. The radiomics model was developed in a cohort of 109 patients from Institution 1, then the model performance was evaluated in the external test cohort of 49 patients from Institutions 2 and 3. The area under the receiver operating characteristic curve (AUC) of the radiomics model was compared with two radiologists using the DeLong test. Hormone testing results were collected to determine the presence of excess cortisol or catecholamines. SHapley Additive explanations (SHAP) was used to improve the interpretability of the radiomics model.
+Yinyao Chao, Hongzhang Zhu, Wenyi Yang and Haohua Yao contributed equally to this work.
*Correspondence: Jingtong Zhang zhangjt@sysucc.org.cn Bin Li binlee@scut.edu.cn Shengjie Guo
Full list of author information is available at the end of the article
☒ BMC
@ The Author(s) 2025. Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creati vecommons.org/licenses/by-nc-nd/4.0/.
Results We developed and evaluated a radiomics model consisting of ten selected CT-based radiomics features. In the external test cohort, the proposed radiomics model achieved high accuracy (86%), specificity (88%), and sensitivity (81%) in differentiating between ACC and PHEO and outperformed 2 radiologists (AUC 0.920 vs. 0.786, 0.629). This radiomics model showed strong capabilities in differentiating biochemically negative ACC and PHEO (with an accuracy of 80%). Moreover, its performance remained consistent even when cortisol and catecholamine levels were simultaneously elevated. Furthermore, SHAP provided quantitative explanations for the radiomics model and visualized the diagnostic process.
Conclusions The interpretable CT-based radiomics model outperforms radiologists in differentiating between ACC and PHEO, especially when hormone testing results are atypical.
Keywords Adrenocortical carcinoma, Pheochromocytoma, Radiomics model, Contrast-enhanced CT, SHapley additive explanations
Background
The past few decades have witnessed the increasing incidence rates of adrenal malignancies [1, 2]. Adrenal malignancies mainly consist of adrenocortical carcinoma (ACC), pheochromocytoma (PHEO), and metastatic tumor of another primary tumor, etc. [3]. ACC is a rare malignancy arising from the adrenal cortex, affecting a mere 1 to 2 individuals per million annually [4]. ACC is characterized by a bleak prognosis, the reported 5-year survival rate ranges between 15 and 44% [5]. On the contrary, PHEO, another adrenal malignancy originat- ing in the adrenal medulla, has a preferable prognosis of 40-95% [6-8].
Accurate diagnosis of ACC is challenging but impor- tant. Currently, morphological and functional evaluation are well-accepted methods for the diagnosis of adrenal tumors. Morphologically, clinical guidelines of Europe or the United States recommend that adrenal tumors with unenhanced computed tomography (CT) densi- ties ≤10 Hounsfield units (HU) and a tumor size<4 cm are likely to be benign, while those with unenhanced CT values> 10 HU or a size≥4 cm have potential of being malignant tumors [3, 9]. However, conventional imag- ing techniques are challenged in differentiating between ACC and PHEO due to their resembling imaging char- acteristics, including similar CT attenuation values, comparable absolute and relative enhancement losses, heterogeneity, and the presence of lipids [10-12].
In terms of functional evaluation, according to the NCCN Guidelines for Neuroendocrine and Adrenal Tumors, adrenal tumors (≥4 cm) with hypercortisolemia and/or Cushing’s syndrome may suggest ACC, whereas those with high levels of catecholamines, such as meta- nephrine and normetanephrine, may suggest PHEO [13]. However, not all patients with ACC have excess cortisols [11, 14], and sometimes ACC patients could have ele- vated catecholamine levels [15]. These atypical functional abnormalities can also occur in patients with PHEO. The presence of either normal catecholamine levels or ele- vated cortisol levels in PHEO patients can complicate the
diagnostic process [16-18]. Thus, the differential diag- nosis between large (24 cm) ACC and PHEO remains a persistent concern and challenge for surgeons and radi- ologists [19-21].
Difficulty in differentiating between large ACC and PHEO with radiology and biochemistry can cause uncer- tainty in selecting surgical approaches for patients, which is crucial. Notably, laparoscopic adrenalectomy instead of traditional open adrenalectomy is the standard approach for localized PHEO, given its preferable perioperative outcomes and low complication rates [22, 23]. Although laparoscopic adrenalectomy also appears to be a viable option for small ACC without evidence of local inva- sion, open adrenalectomy is still the standard procedure for ACC patients, for its clearer advantage in prevent- ing tumor capsule breaching and local dissemination [11, 13, 24]. In addition, preoperative administration of alpha-adrenergic blockers, such as phenoxybenzamine, is essential in managing PHEO. This approach aims to expand blood volume, stabilize blood pressure and heart rate, and minimize intraoperative cardiovascular compli- cations [25]. Therefore, accurate differentiation between large ACC and PHEO is crucial for guiding surgical deci- sions and improving prognosis.
Radiomics offers an opportunity to classify adre- nal tumors. However, previous research has primarily focused on differentiating malignant adrenal lesions from benign cases, including the differentiation of adrenocor- tical adenoma (ACA) from ACC [26, 27], lipid-poor ACA from malignant adrenal tumors [28], as well as PHEO from other adrenal lesions [29-31]. In this case, there has been conspicuously limited research on the differen- tiation between ACC and PHEO. Moreover, the rarity of ACC makes it difficult to initiate large-scale studies, thus hindering further applications of radiomics.
In this study, we aim to develop an interpre- table radiomics model for the preoperative differ- entiation between large ACC and PHEO based on contrast-enhanced CT and subsequently assess the
model’s performance against the professional judgment of radiologists and hormone testing results.
Methods
Study participant
The Institutional Ethics Committee of Sun Yat-sen Uni- versity Cancer Center approved this retrospective study in accordance with the Declaration of Helsinki, and the need for informed consent was waived (SL-B2024-545- 01). The CheckList for EvaluAtion of Radiomics research (CLEAR) and METhodological RadiomICs Score (MET- RICS) were attached in the Supplementary CLEAR Checklist and Supplementary METRICS Checklist, respectively [32, 33].
Between January 2011 and September 2023, 541 con- secutive patients from three independent medical insti- tutions (Institution 1: Sun Yat-sen University Cancer Center, China; Institution 2: The First Affiliated Hos- pital, Sun Yat-sen University, China; Institution 3: Nan- fang Hospital, Southern Medical University, China), who underwent either adrenalectomy or adrenal biopsy and then received pathological diagnoses of ACC or PHEO, were retrospectively recruited. Exclusion criteria were as follows: (a) contrast-enhanced CT imaging quality was unsatisfying, (b) contrast-enhanced images were incom- plete, (c) tumor size was less than 4 cm, (d) tumor was either recurrent or metastatic.
A total of 158 patients (median age, 47 years; inter- quartile range, 32-57 years; 76 males) were enrolled in this study, with 109 patients from Institution 1 as the model development cohort and 49 patients from Institu- tion 2 and 3 as the external test cohort (Fig. 1A).
CT technique
Contrast-enhanced CT scans of the upper abdomen, including bilateral adrenal glands and lesions, were per- formed following a standardized scanning protocol at three institutions. Imaging protocols and CT systems were detailed in Supplementary Material 1.
Image acquisition and segmentation
The radiomics workflow was shown in Fig. 1B. Briefly, all CT images were loaded into the 3D Slicer (version 5.0.3, https://www.slicer.org) to undergo segmentation. All images’ regions of interest (ROIs) were manually seg- mented slice-by-slice by reader 1 (J. Zhao, a senior radi- ologist at Institution 1, who had more than ten years of expertise in urinary system tumors). The segmentation was validated by reader 2 (H. Ma, a junior radiologist at Institution 1, who had more than five years of exper- tise in urinary system tumors) to assess the repeatability in a randomly selected cohort of 20 patients (six ACC cases and fourteen PHEO cases) from the development
cohort. Both readers were blinded to patients’ pathologi- cal diagnoses.
Image preprocessing and radiomics feature extraction
Following the Imaging Biomarker Standardized Initia- tive (IBSI) guidelines, CT images were resampled to a voxel size of 1 × 1 × 1 mm to eliminate the effect of differ- ent slice intervals between the institutions [34]. Prior to feature extraction, we processed the CT images using the Laplacian of Gaussian (LoG) and wavelet filters to achieve multiple sets of radiomics features. We used Pyradiomics (version 3.0.1, https://github.com/Radiomics/pyradio mics), an open-sourced package to extract radiomics fe atures from non-enhanced, arterial phase, and venous phase images. The radiomics features extracted from the images can be categorized into the following classes: first order statistics, shaped-based features, gray level co- occurrence matrix (GLCM), gray level run length matrix (GLRLM), gray level size zone matrix (GLSZM), neigh- boring gray tone difference matrix (NGTDM), and gray level dependence matrix (GLDM). After feature extrac- tion, we applied ComBat (https://github.com/Jfortin1/C omBatHarmonization), a batch-effect correction tool to compensate for multicenter effects among different insti- tutions, and additionally employed z-score to normalize data across different phases [35].
Radiomics feature selection
We followed a two-step procedure to identify radiomics features used in model training. Initially, we used the interclass correlation coefficient (ICC) to evaluate inter- rater reliability, and features with values over 0.75 were preserved [36]. In this study, we applied SHapley Addi- tive explanations with Recursive Feature Elimination with Cross-validation (SHAP-RFECV) from the Pro- batus library (https://ing-bank.github.io/probatus) to perform feature selection for the XGBoost algorithm [37]. In the feature elimination process, we used a ran- domized hyperparameter search schema from the sci-kit learn library to obtain the optimal hyperparameters that achieve the highest mean area under the receiver operat- ing characteristic curve (AUC) in ten-fold cross-valida- tion of the development cohort.
Model development and evaluation
Based on the optimal feature set and hyperparam- eters in the feature selection process, we performed a ten-fold cross-validation on the development cohort to further confirm its performance. Subsequently, we retrained the XGBoost using the entire development cohort to obtain a final radiomics model for differenti- ating between ACC and PHEO. Performance evaluation of the radiomics model was carried out on the external test cohort using metrics including AUC, the area under
A
Patients who underwent either adrenalectomy or adrenal biopsy with pathological diagnosis of ACC or PHEO from 2011 to 2023 at Institution 1 (n = 263)
Patients who underwent either adrenalectomy or adrenal biopsy with pathological diagnosis of ACC or PHEO from 2011 to 2023 at Institution 2 (n = 193)
Patients who underwent either adrenalectomy or adrenal biopsy with pathological diagnosis of ACC or PHEO from 2011 to 2023 at Institution 3 (n = 85)
154 patients excluded:
(a) contrast-enhanced CT imaging quality was unsatisfying (n = 5)
163 patients excluded:
66 patients excluded:
(b) contrast-enhanced CT images were incomplete (n = 4)
(a) contrast-enhanced CT imaging quality was unsatisfying (n = 2)
(a) contrast-enhanced CT imaging quality was unsatisfying (n = 4)
(c) tumor size was less than 4 cm (n = 129) (d) tumor was either recurrent or metastatic (n = 6)
(b) contrast-enhanced CT images were incomplete (n = 149)
(b) contrast-enhanced CT images were incomplete (n = 4)
(c) tumor size was less than 4 cm (n = 11)
(d) tumor was either recurrent or metastatic (n = 1)
(c) tumor size was less than 4 cm (n = 58)
(d) tumor was either recurrent or metastatic (n = 0)
The development cohort (n = 109) ACC (n = 37), PHEO (n = 72)
Institution 2 (n = 30) ACC (n = 13), PHEO (n = 17)
Institution 3 (n = 19) ACC (n = 3), PHEO (n = 16)
The test cohort (n = 49) ACC (n = 16), PHEO (n = 33)
Model construction
B
LoG filter
LoG-based features
Resampling
Original features
Wavelet-based features
Original image
Wavelet filter
Adrenal tumor Image acquisition
ROIs segmentation
Image preprocessing and feature extraction
Fold 1
ICC > 0.75
SHAP summary plot
Fold 2
Training
Fold 3
Development cohort
1.0
Model AUC performance
…
Validation
0.8
Fold 10
-0.6
Waterfall plot of patients
10-fold cross-validation
AUC = 0.918 Train score (SD) Validation score (SD)
0.4
0.2
Test cohort
Development cohort
0 10 21 32
48 60
74
92
4004
0.0
Number of features
Model interpretability
Model development and evaluation
Feature selection
the precision-recall curve (PRAUC), accuracy, sensitivity (TPR), specificity (TNR), and F1 score. In addition, we used SHapley Additive explanations (SHAP) to provide an intuitive visualization of the importance of features, in order to help us understand the impact of each feature on the model predictions [38].
To assess the diagnostic efficacy of radiologists, read- ers 1 and 2 reviewed all CT images from the external test cohort independently. They were aware of patients’ age, sex, and the existence of adrenal abnormalities but
blinded to pathological diagnoses. They had strictly fol- lowed the criteria outlined in the guidelines from the European Society of Endocrinology, NCCN, and the expert consensus in China when analyzing the CT images [3, 13, 39, 40]. AUC and other metrics of radiologists’ diagnoses were compared with the radiomics model.
We collected patients’ preoperative hormone testing results from hospital information systems. Excess corti- sols was defined by the presence of at least one measure- ment of cortisol hormone or its metabolites exceeding
| Development cohort | External test cohort | p value | ||
|---|---|---|---|---|
| Number | 109 | 49 | ||
| Sex | Male | 56 (51) | 20 (41) | |
| Female | 53 (49) | 29 (59) | 0.22 | |
| Age (y) | 46 (32, 55) | 49 (34, 61) | 0.40 | |
| BMI | 21.9 (19.5, | 21.6 (19.7, 23.9) | 0.88 | |
| 24.2) | ||||
| Tumor size | 70 (56, 96.5) | 70 (60, 96.3) | 0.77 | |
| (mm) |
Note. Except where indicated, data were numbers of patients, with percentages in parentheses. Age, BMI, and tumor size were presented as medians with inter- quartile ranges in parentheses. Mann-Whitney U test or x2 test was used to calculate the p value. BMI = Body mass index
the upper limit of the reference range, including plasma cortisol, 24-hour urinary 17-ketosteroids, 24-hour uri- nary 17-hydroxysteroid, and 24-hour urinary free cor- tisol. Similarly, excess, normal or low catecholamines were also defined based on biochemistry testing of cat- echolamines, including blood or urinary vanillylman- delic acid (VMA), the ratio of VMA to creatinine, the ratio of homovanillic acid (HVA) to creatinine, epineph- rine, norepinephrine, metanephrine, normetanephrine, 3-methoxytyramine [17].
Statistical analysis
Categorical and continuous variables of patients’ char- acteristics were compared using the x2 test and Mann- Whitney U test, respectively. Given the expected AUC of 0.92, the prevalence in the evaluation population of 0.33, and the AUC’s target width of 0.2, the minimum required sample size of patients in the external test cohort was 37 [41]. MedCalc (version 22.019, https://www.medcalc.org) was applied for all statistical analyses. Receiver operat- ing characteristic (ROC), precision-recall curve (PRC), and confusion matrix were used to assess diagnostic
performance. ROCs of radiologists and the radiomics model were compared using the DeLong test, while accu- racies were compared using the McNemar test. Post hoc power analysis was performed at https://www.medcalc .org/calc/post-hoc-power-analysis.php. We considered p <0.05 to be statistically significant.
Results
Patient characteristic
Patients’ baseline clinical features in the development cohort (Institution 1, n = 109, ACC: 37 cases, PHEO: 72 cases) and the external test cohort (Institution 2 and Institution 3, n = 49, ACC: 16 cases, PHEO: 33 cases) were shown in Table 1. There was no significant difference in the ratio of ACC to PHEO (p=0.87), and the tumor size (inter-quartile range of the development cohort, 56-97; inter-quartile range of the external test cohort, 60-96.5, p=0.77) between the development cohort and the exter- nal test cohort.
Model building and performance
4227 (1409 of each phase) radiomics features, including the original features, the LoG-based features, and the wavelet-based features, were reduced to 4004 in the first step of feature selection (ICC>0.75). Figure 2 showed that the optimal feature set with the optimized hyper- parameters achieved the highest mean AUC of 0.918 with ten radiomics features in the development cohort. The ten most relevant features selected were three origi- nal, three LoG-based, and four wavelet-based features, among which there were two features from the non- enhanced phase, two from the arterial phase, and five from the venous phase, respectively (Table 2).
In the ten-fold cross-validation of the development cohort, the radiomics model yielded high AUC (mean, 0.918; standard deviation, 0.141), accuracy (mean, 0.936;
Development cohort
1.0
Model AUC performance
0.8
0.6
AUC = 0.918
0.4
Train score (SD)
Validation score (SD)
0.2
0
10
0.0
21
32
48
60
74
92
4004
Number of features
| Filter | Contrast-enhanced CT phases | ||
|---|---|---|---|
| Non-enhanced phase | Arterial phase | Venous phase | |
| Original images | Original-glszm-SizeZoneNonUniformityNormalized | Original-ngtdm-Contrast Original-glcm-ClusterShade | |
| Laplacian of Log-sigma-1-0-mm-3D-glcm-MaximumProbability Gaussian | Log-sigma-1-5-mm-3D-glcm-MCC | Log-sigma-1-0-mm-3D- gldm-DependenceNon- Uniformity | |
| Wavelet Wavelet-LHH-glszm-LargeAreaHighGrayLevelEm- phasis Wavelet-HLH-glcm-Autocorrelation | Wavelet-HLL-glszm-SmallA- reaLowGrayLevelEmphasis Wavelet-LLL-glszm-Low- GrayLevelZoneEmphasis | ||
| Folds | Accuracy | Sensitivity | Specificity | AUC |
|---|---|---|---|---|
| 1 | 11/11 (1.000) | 4/4 (1.000) | 7/7 (1.000) | 1.000 |
| 2 | 8/11 (0.727) | 1/4 (0.250) | 7/7 (1.000) | 0.571 |
| 3 | 10/11 (0.909) | 4/4 (1.000) | 6/7 (0.857) | 0.929 |
| 4 | 10/11 (0.909) | 4/4 (1.000) | 6/7 (0.857) | 0.893 |
| 5 | 9/11 (0.818) | 4/4 (1.000) | 5/7 (0.714) | 0.786 |
| 6 | 11/11 (1.000) | 4/4 (1.000) | 7/7 (1.000) | 1.000 |
| 7 | 11/11 (1.000) | 4/4 (1.000) | 7/7 (1.000) | 1.000 |
| 8 | 11/11 (1.000) | 3/3 (1.000) | 8/8 (1.000) | 1.000 |
| 9 | 11/11 (1.000) | 3/3 (1.000) | 8/8 (1.000) | 1.000 |
| 10 | 11/11 (1.000) | 3/3 (1.000) | 7/7 (1.000) | 1.000 |
Note. AUC=area under the receiver operating characteristic curve
standard deviation, 0.096), specificity (mean, 0.943; stan- dard deviation, 0.100), and sensitivity (mean, 0.925; stan- dard deviation, 0.237). Details of cross-validation were provided in Table 3.
In the external test cohort, the radiomics model achieved a high AUC of 0.920 (95% CI: 0.845, 0.996, Fig. 3A). And Fig. 3B showed the precision-recall curve. The accuracy, specificity, and sensitivity of the exter- nal test cohort were 0.857 (95% CI: 0.728, 0.941), 0.879
| Evaluation metric | Radiomics model | Reader 1 | Reader 2 |
|---|---|---|---|
| AUC | 0.920 | 0.786 | 0.629 |
| [0.845, 0.996] | [0.645, 0.890] | [0.479, 0.762] | |
| PRAUC | 0.871 | 0.564 | 0.441 |
| [0.610, 0.967] | [0.325, 0.776] | [0.227,0.679] | |
| Accuracy | 0.857 (42/49) | 0.755 (37/49) | 0.673 (33/49) |
| [0.728, 0.941] | [0.611, 0.866] | [0.525, 0.801] | |
| Sensitivity | 0.813 (13/16) | 0.875 (14/16) | 0.500 (8/16) |
| [0.544, 0.960] | [0.617, 0.984] | [0.247, 0.753] | |
| Specificity | 0.879 (29/33) | 0.697 (23/33) | 0.758 (25/49) |
| [0.718, 0.966] | [0.513, 0.844] | [0.577,0.889] | |
| F1 max | 0.800 | 0.700 | 0.500 |
Note. Data in parentheses are numbers of patients, with 95% confidence interval in brackets. AUC=area under the receiver operating characteristic curve. PRAUC =area under the precision-recall curve
(95% CI: 0.718, 0.966), and 0.813 (95% CI: 0.544, 0.960), respectively (Table 4). The prediction values of the radiomics model ranged from 0 to 1. Higher prediction values (closer to 1) indicated ACC, while lower values (closer to 0) suggested PHEO, with a cut-off value of 0.305. All prediction values and the calibration curve of
A
External test cohort
B
External test cohort
1.0-
1.0
0.8.
0.8-
Sensitivity (TPR)
0.6
Precision
0.6
0.4-
0.4
AUC = 0.920 [0.845, 0.996]
PRAUC = 0.871 [0.610, 0.967]
0.2-
Reader 1
0.2-
Reader 1
Reader 2
Reader 2
0.0
0.0
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
1-Specificity (FPR)
Recall
A
Radiomics model
B
Reader 1
C
30
30
Reader 2
30
1
3
13
-
20
2
14
-
20
8
8
20
True
True
True
10
10
10
0
29
4
0
23
10
0
25
8
0
0
1
0
-0
Predicted
0
Predicted
1
0
Predicted
1
| Development cohort | External test cohort | |||
|---|---|---|---|---|
| ACC | PHEO | ACC | PHEO | |
| Excess cortisols | 39.3% | 29.7% | 66.7% | 18.8% |
| (11/28) | (19/64) | (10/15) | (6/32) | |
| Normal or low cortisols | 60.7% | 70.3% | 33.3% | 81.2% |
| (17/28) | (45/64) | (5/15) | (26/32) | |
| Excess catecholamines | 5.9% (1/17) | 58.0% | 0.0% | 71.0% |
| (29/50) | (0/9) | (22/31) | ||
| Normal or low | 94.1% | 42.0% | 100% (9/9) | 29.0% |
| catecholamines | (16/17) | (21/50) | (9/31) | |
| Excess cortisols and | 5.9% (1/17) | 10.2% | 0.0% | 16.7% |
| excess catecholamines | (5/49) | (0/9) | (5/30) | |
| Neither excess | 47.1% (8/17) | 6.1% | 33.3% (3/9) | 23.3% |
| cortisols nor excess | (3/49) | (7/30) | ||
| catecholamines | ||||
Note. Data in parentheses are numbers of patients
the external test cohort were provided in Supplementary Material 2.
Comparison of the radiomics model with radiologists
In the external test cohort, reader 1 achieved an AUC of 0.786 (95% CI: 0.645, 0.890), an accuracy of 0.755 (95% CI: 0.611, 0.866), a sensitivity of 0.875 (95% CI: 0.617, 0.984), and a specificity of 0.697 (95% CI: 0.513, 0.844). Reader 2 also achieved an AUC of 0.629 (95% CI:0.479, 0.762), an accuracy of 0.673 (95% CI: 0.525, 0.801), a sen- sitivity of 0.500 (95% CI: 0.247, 0.753), and a specificity of 0.758 (95% CI: 0.577, 0.889) (Table 4). Details regarding the radiologists’ diagnoses can be found in Supplemen- tary Material 3.
Notably, the radiomics model achieved high diagnos- tic accuracy. In terms of AUC, the radiomics model out- performed reader 1 (0.920 vs. 0.786, p<0.05) and reader 2 (0.920 vs. 0.629, p<0.01). There was no significant dif- ference in AUC between reader 1 and reader 2 (0.786 vs. 0.629, p=0.13). Besides, there were no significant dif- ferences in accuracy between the radiomics model and reader 1 (p=0.12), the radiomics model and reader 2 (p=1.00), as well as reader 1 and reader 2 (p=0.13).
Due to the relatively small sample size, we performed a post hoc power analysis. The analysis revealed a power of >0.99 for reader 1 and 0.579 for reader 2, with a p of 0.05. The diagnostic confusion matrixes were shown in Fig. 4.
Application of the radiomics model in cases with atypical hormone levels
Of all 158 patients enrolled in this study, 90% had under- gone testing for cortisol or catecholamine levels. Surpris- ingly, 49% of ACC patients did not have elevated cortisol levels, and 37% of PHEO patients did not have elevated catecholamine levels (Table 5), which was unexpected (Supplementary Material 4).
Notably, in the external test cohort, the radiomics model demonstrated strong diagnostic capability in identifying biochemically atypical ACC and PHEO. The model successfully identified 80% of ACC cases that did not have excess cortisols and 78% of PHEO cases that did not have excess catecholamines (Supplementary Material 5).
Additionally, the radiomics model correctly diagnosed all six PHEO patients who exhibited both cortisol and catecholamine excess, and correctly identified 90% of patients (three with ACC and seven with PHEO) who show neither excess cortisols nor excess catecholamines. In these cases, the radiomics model achieved an accuracy of 0.938 (95% CI: 0.698, 0.998), a sensitivity of 1.000 (95% CI: 0.292, 1.000), and a specificity of 0.923 (95% CI: 0.640, 0.998). Details were provided in Supplementary Material 6.
Interpretability of the radiomics model
The interpretability of the radiomics model was improved using SHAP. Through the SHAP summary plot, a clear visual representation of ten selected features was pre- sented, showing how the values of these features corre- lated with their impact on the model output (Fig. 5A). The features were ranked based on their relevance to model output, and the “original-glcm-ClusterShade” from the venous phase was identified as the most rel- evant feature.
A
module-3-original-glcm-ClusterShade
wavelet-LHH-glszm-LargeAreaHighGrayLevelEmphasis
High
module-3-wavelet-HLL-glszm-SmallAreaLowGrayLevelEmphasis
log-sigma-1-0-mm-3D-glcm-MaximumProbability
Feature value
module-3-wavelet-LLL-glszm-LowGrayLevelZoneEmphasis
module-3-log-sigma-1-0-mm-3D-gldm-DependenceNonUniformity
module-2-original-glszm-SizeZoneNonUniformityNormalized
module-3-original-ngtdm-Contrast
module-2-log-sigma-1-5-mm-3D-glcm-MCC
Low
wavelet-HLH-glcm-Autocorrelation
-0.15
-0.10
-0.05
0.00
0.05
0.10
0.15
0.20
0.25
SHAP value (impact on model output)
B
Non-enhanced phase
Arterial phase
Venous phase
Tumor
Tumor
Tumor
C
f(x)
: 0.570
-0.427 = module-3-original-glcm-ClusterShade
+0.218
1.191 = log-sigma-1-0-mm-3D-glcm-MaximumProbability
+0.099
-1.08 = wavelet-LHH-glszm-LargeAreaHighGrayLevelEmphasis
-0.065
-0.578 = module-3-wavelet-HLL-glszm-SmallAreaLowGrayLevelEmphasis-
+0.047
-0.365 = module-3-wavelet-LLL-glszm-LowGrayLevelZoneEmphasis
+0.031
-0.568 = wavelet-HLH-glcm-Autocorrelation ·
-0.028
-0.193 = module-3-original-ngtdm-Contrast -
-0.028
-0.615 = module-2-log-sigma-1-5-mm-3D-glcm-MCC
-0.022
-0.161 = module-2-original-glszm-SizeZoneNonUniformityNormalized
-0.008
-0.143 = module-3-log-sigma-1-0-mm-3D-gldm-DependenceNonUniformity
-0.002
0.20
0.25
0.30
0.35
0.40
0.45
0.50
0.55
E[f(x)] = 0.328
0.60
The potential application of this model was highlighted by a case in Institution 3 (patient No. 31) who was pre- dicted to have ACC by our radiomics model (Fig. 5B). This patient initially presented with a 6 cm adrenal
incidentaloma and was given a radiological diagnosis of PHEO. Laparoscopic adrenalectomy was performed by urologists, but post-surgery pathology revealed the lesion was ACC with capsular invasion. Regrettably, the patient
developed multiple lymph nodes, pulmonary, and hepatic metastases within 8 months following the laparoscopic procedure.
In the retrospective analysis, reader 1 correctly pre- dicted the lesions as ACC, whereas reader 2 gave a PHEO diagnosis. The radiomics model prediction value of this patient was 0.570, which exceeded the cut-off value of 0.305, therefore the model diagnosed the lesion as ACC. A waterfall plot was shown to illustrate how the model outputted an individual prediction of this patient based on different radiomics features (Fig. 5C). In all, this case strongly demonstrated the radiomics model’s advantage in predicting ACC from highly deceptive CT images.
Discussion
Adrenocortical carcinoma (ACC), a rare and highly malignant adrenal gland tumor, exhibits CT character- istics that resemble those of the less malignant pheo- chromocytoma (PHEO). In addition, while biochemical evaluation is widely accepted for differentiating between ACC and PHEO in some cases, non-functioning tumors remain a diagnostic challenge [20, 21]. This dilemma can lead to suboptimal accuracy in diagnosis, increasing the risk of misdiagnosis or inappropriate clinical interven- tions, potentially leading to serious consequences.
Here, in our multicenter study, we developed and tested a radiomics model that showed promising diagnostic per- formance in differentiating between large (24 cm) ACC and PHEO in both internal and external cohorts. In the development cohort, the radiomics model showed high AUC, accuracy, specificity, and sensitivity (0.918, 0.936, 0.943, and 0.925). In the external test cohort, our model showed substantial improvement over radiologists, achieving a remarkable AUC of 0.920, accuracy of 0.857, specificity of 0.879, and sensitivity of 0.813. Notably, the radiomics model also showed potential clinical utility in cases which hormone testing results were atypical.
Contrast-enhanced CT is a widely recommended imag- ing modality to localize and evaluate adrenal tumors [42]. Clinical guidelines primarily consider adrenal tumors with the unenhanced CT values ≤ 10 HU and a size <4 cm to be benign. Tumors that do not meet any of the crite- ria may be malignant, and some PHEO and ACC fall into this category [3, 9]. In addition, PHEO and ACC often show similar CT imaging features, such as heterogeneous and lipid-containing, etc., which can lead to suboptimal accuracy in CT diagnosis, highlighting the imperative for improved diagnostic approaches [4, 12, 19].
Hormone testing can assist in differentiating between ACC and PHEO, but the results may occasionally be atypical. Clinical guidelines from European Society of Endocrinology, NCCN, and ESMO recommend using fractionated metanephrines for diagnosing PHEO [3, 13, 24]. However, some patients with ACC may also have
elevated levels of metanephrines, leading to false positive results and diagnostic challenges [11, 15, 43]. Further- more, a study shows that patients with PHEO can exhibit elevated plasma cortisol levels (inter-quartile range, 259- 452 nmol/L) compared to a reference group, with some overlap with Cushing syndrome patients (inter-quartile range, 390-636 nmol/L). This issue undoubtedly made the diagnosis of ACC and PHEO more challenging and our data further substantiated this situation [18].
Radiomics, a method of converting medical images into higher-dimensional data, has the potential to be applied to the diagnosis of adrenal malignancy [44]. However, the existing radiomics studies mainly focus on the dif- ferentiation of malignancies from benign adrenal tumors, with limited attention paid to the specific classification of malignancies such as ACC and PHEO [29, 30, 45]. Previ- ous radiomics studies of ACC mainly focus on the differ- entiation between ACC and ACA, and those studies are often hampered by the rarity of ACC, therefore no multi- center study has been performed to date [26, 28, 46]. For instance, Elmohr et al. performs a CT texture analysis on 54 patients (25 with ACA, 29 with ACC, all >4 cm), and their model yields an AUC of 0.886 [46].
The optimal feature set of this radiomics model con- tained ten features extracted from three CT phases, sug- gesting that the integration of features from multiple phases can improve diagnostic performance. Interest- ingly, half of these features are those from the venous phase. Our finding, together with clinical recommenda- tion, underscores the significance of the venous phase imaging in differentiating between ACC and PHEO [42].
In radiomics studies, machine learning-based algo- rithms extract higher-dimensional data. However, the specific mechanisms behind the radiomics models are often opaque and might limit models’ acceptance in clinical practice [47, 48]. In this study, the applica- tion of SHAP improved the overall interpretability of the radiomics model. It demonstrated the impact of radiomics features on model prediction and provided quantitative explanations of each patient’s predictive pro- cess. Notably, the radiomics model accurately predicted a case with ACC, who was preoperatively diagnosed with PHEO and underwent laparoscopic adrenalectomy. In this case, the predictive process was visualized by SHAP, which can help clinicians better understand the predic- tive capabilities of the radiomics model.
To our knowledge, this is the first multicenter study to show that a CT-based radiomics model can be used to differentiate between large ACC and PHEO preopera- tively. This radiomics model was based on open-sourced programs. It not only showed excellent performance across the multicenter cohort, but also was equipped with improved interpretability through the application of SHAP.
However, there are several limitations to be acknowl- edged. First, it is a retrospective study and selection bias should be considered, despite we performed exter- nal validation to improve reliability. Second, the study is designed using a modest number of patients due to the rarity of ACC. Therefore, a prospective study with larger patient samples is needed to define the risk of selection bias and improve diagnostic accuracy.
In conclusion, we developed a novel interpretable radiomics model based on contrast-enhanced CT for differentiating between large ACC and PHEO preop- eratively. Our radiomics model stands out for its inter- pretability, powerfulness, and generalizability, providing radiologists and surgeons a valuable tool to optimize the diagnostic workflow for adrenal malignancies. With fur- ther prospective validations and refinements, it can offer an accurate and reliable approach to improving clinical decision-making.
Conclusions
The interpretable radiomics model, based on contrast- enhanced CT, outperforms radiologists in differentiating between ACC and PHEO. This model can serve as a valu- able tool to optimize the diagnostic accuracy for adrenal malignancies, especially when hormone testing results are atypical.
Abbreviations
| ACC | Adrenocortical carcinoma |
| PHEO | Pheochromocytoma |
| CT | Computed tomography |
| HU | Hounsfield units |
| ACA | Adrenocortical adenoma |
| ROI | Region of interest |
| LoG | Laplacian of Gaussian |
| ICC | Interclass correlation coefficient |
| AUC | Area under the receiver operating characteristic curve |
| PRAUC | The area under the precision-recall curve |
| SHAP | SHapley Additive explanations |
| VMA | Vanillylmandelic acid |
| HVA | Homovanillic acid |
| ROC | Receiver operating characteristic |
| CI | Confidence interval |
Supplementary Information
The online version contains supplementary material available at https://doi.or g/10.1186/s12880-025-01842-7.
Supplementary Materials
Acknowledgements
Not applicable.
Author contributions
S. G, B. L, and J. Z conceived, designed, and directed the study. Y. C, H. Z, W. Y, and H. Y acquired the data, formed the analyses and interpretation of data. Y. C, W. Y, and H. Y developed the algorithms. J. Z, Y. C, H. Z, and H. Y prepared figures. N. M, X. C, Z. L, H. H, Z. L, and K. Y provided technical support. J. Z and H. M performed image segmentation. H. Z, Y. L, and P. W provided multicenter data. S. G, B. L, and J. Z wrote and critically revised the manuscript. All authors
read and approved the final manuscript. The order of the co-first authors was assigned based on the relative contributions of the individuals.
Funding
This study was supported by the Guangdong Basic and Applied Basic Research Foundation (No. 2023A1515110515, No. 2022A1515111119) and the Fostering Program for NSFC Young Applicants (Tulip Talent Training Program) of Sun Yat-sen University Cancer Center (No. 2025yfd11).
Data availability
The codes of the feature extraction and radiomics model are available at htt ps://github.com/Chaoyinyao/Radiomics-model-to-differentiate-between-a drenocortical-carcinoma-and-pheochromocytoma. The data that support the findings of this study are available based on reasonable request from the corresponding author Shengjie Guo.
Declarations
Ethics approval and consent to participate
The Institutional Ethics Committee of Sun Yat-sen University Cancer Center approved this retrospective study in accordance with the Declaration of Helsinki, and the need for informed consent was waived (SL-B2024-545-01).
Consent for publication
Not applicable.
Competing interests
The authors declare no competing interests.
Author details
1Department of Urology, Sun Yat-sen University Cancer Center, Sun Yat- sen University, Guangzhou 510060, China
2State Key Laboratory of Oncology in Southern China, Collaborative
Innovation Center of Cancer Medicine, Guangzhou, China
3Guangdong Provincial Clinical Research Center for Cancer, Guangzhou, China
4Zhongshan School of Medicine, Sun Yat-sen University, Guangzhou, China
5Department of Radiology, The First Affiliated Hospital, Sun Yat-sen University, Guangzhou, China
6School of Automation Science and Engineering, South China University of Technology, Guangzhou 510640, China
7Department of Urology, Guangdong Provincial People’s Hospital, Guangzhou, China
8Department of Radiology, Sun Yat-sen University Cancer Center, Sun Yat- sen University, Guangzhou, China
9Department of Urology, Nanfang Hospital, Southern Medical University, Guangzhou, China
10Department of Urology, Sun Yat-sen University Cancer Center Gansu
Hospital, Sun Yat-sen University, Lanzhou, China
Received: 15 May 2025 / Accepted: 22 July 2025
Published online: 01 August 2025
References
1. . Chandrasekar T, Goldberg H, Klaassen Z, Wallis CJD, Woon DTS, Herrera- Caceres JO, et al. The who, when, and why of primary adrenal malig- nancies: insights into the epidemiology of a rare clinical entity. Cancer. 2019;125(7):1050-9.
2. Lubomski A, Falhammar H, Torpy DJ, Rushworth RL. The epidemiology of primary and secondary adrenal malignancies and associated adrenal insuf- ficiency in hospitalised patients: an analysis of hospital admission data, NSW, Australia. BMC Endocr Disord. 2021;21(1):141.
3. Fassnacht M, Tsagarakis S, Terzolo M, Tabarin A, Sahdev A, Newell-Price J, et al. European society of endocrinology clinical practice guidelines on the management of adrenal incidentalomas, in collaboration with the European network for the study of adrenal tumors. Eur J Endocrinol. 2023;189(1):G1-42.
4. Else T, Kim AC, Sabolch A, Raymond VM, Kandathil A, Caoili EM, et al. Adreno- cortical carcinoma. Endocr Rev. 2014;35(2):282-326.
5. Ayala-Ramirez M, Jasim S, Feng L, Ejaz S, Deniz F, Busaidy N, et al. Adrenocorti- cal carcinoma: clinical outcomes and prognosis of 330 patients at a tertiary care center. Eur J Endocrinol. 2013;169(6):891-9.
6. Lam AK. Update on adrenal tumours in 2017 world health organization (WHO) of endocrine tumours. Endocr Pathol. 2017;28(3):213-27.
7. Juhlin CC, Mete O. Advances in adrenal and extra-adrenal paraganglioma: practical synopsis for pathologists. Adv Anat Pathol. 2023;30(1):47-57.
8. Moog S, Castinetti F, DoCao C, Amar L, Hadoux J, Lussey-Lepoutre C, et al. Recurrence-free survival analysis in locally advanced pheochromocytoma: first appraisal. J Clin Endocrinol Metab. 2021;106(9):2726-37.
9. Zeiger MA, Thompson GB, Duh QY, Hamrahian AH, Angelos P, Elaraj D, et al. The American association of clinical endocrinologists and American associa- tion of endocrine surgeons medical guidelines for the management of adrenal incidentalomas. Endocr Pract. 2009;15(Suppl 1):1-20.
10. Szolar DH, Korobkin M, Reittner P, Berghold A, Bauernhofer T, Trummer H, et al. Adrenocortical carcinomas and adrenal pheochromocytomas: mass and enhancement loss evaluation at delayed contrast-enhanced CT. Radiology. 2005;234(2):479-85.
11. Fassnacht M, Dekkers OM, Else T, Baudin E, Berruti A, de Krijger R, et al. European society of endocrinology clinical practice guidelines on the management of adrenocortical carcinoma in adults, in collaboration with the European network for the study of adrenal tumors. Eur J Endocrinol. 2018;179(4):G1-46.
12. Mendiratta-Lala M, Avram A, Turcu AF, Dunnick NR. Adrenal imaging. Endocri- nol Metab Clin North Am. 2017;46(3):741-59.
13. National Comprehensive Cancer Network. Neuroendocrine and Adrenal Tumors (Version 1.2024). https://www.nccn.org/professionals/physician_gls/ pdf/neuroendocrine.pdf. Accessed 16 July 2024.
14. Vanbrabant T, Fassnacht M, Assie G, Dekkers OM. Influence of hormonal functional status on survival in adrenocortical carcinoma: systematic review and meta-analysis. Eur J Endocrinol. 2018;179(6):429-36.
15. Kannan S, Purysko A, Faiman C, Remer EM, Shah L, Bena J, et al. Biochemical and radiological relationships in patients with pheochromocytoma: lessons from a case control study. Clin Endocrinol (Oxf). 2014;80(6):790-6.
16. Constantinescu G, Preda C, Constantinescu V, Siepmann T, Bornstein SR, Lenders JWM, et al. Silent pheochromocytoma and paraganglioma: system- atic review and proposed definitions for standardized terminology. Front Endocrinol (Lausanne). 2022;13:1021420.
17. Eisenhofer G, Pamporaki C, Lenders JWM. Biochemical assessment of pheo- chromocytoma and paraganglioma. Endocr Rev. 2023;44(5):862-909.
18. Constantinescu G, Langton K, Conrad C, Amar L, Assié G, Gimenez-Roqueplo AP, et al. Glucocorticoid excess in patients with pheochromocytoma compared with paraganglioma and other forms of hypertension. J Clin Endocrinol Metab. 2020;105(9):e3374-83.
19. Blake MA, Cronin CG, Boland GW. Adrenal imaging. AJR Am J Roentgenol. 2010;194(6):1450-60.
20. Ahmed AA, Thomas AJ, Ganeshan DM, Blair KJ, Lall C, Lee JT, et al. Adrenal cortical carcinoma: pathology, genomics, prognosis, imaging features, and mimics with impact on management. Abdom Radiol (NY). 2020;45(4):945-63.
21. Bharwani N, Rockall AG, Sahdev A, Gueorguiev M, Drake W, Grossman AB, et al. Adrenocortical carcinoma: the range of appearances on CT and MRI. AJR Am J Roentgenol. 2011;196(6):W706-14.
22. Neumann HPH, Young WF Jr., Eng C. Pheochromocytoma and paragan- glioma. N Engl J Med. 2019;381(6):552-65.
23. Li J, Wang Y, Chang X, Han Z. Laparoscopic adrenalectomy (LA) vs open adrenalectomy (OA) for pheochromocytoma (PHEO): a systematic review and meta-analysis. Eur J Surg Oncol. 2020;46(6):991-8.
24. Fassnacht M, Assie G, Baudin E, Eisenhofer G, de la Fouchardiere C, Haak HR, et al. Adrenocortical carcinomas and malignant phaeochromocytomas: ESMO-EURACAN clinical practice guidelines for diagnosis, treatment and follow-up. Ann Oncol. 2020;31(11):1476-90.
25. Friedman LR, Ramamoorthy B, Nilubol N. Progress in surgical approaches and outcomes of patients with pheochromocytoma and paraganglioma. Best Pract Res Clin Endocrinol Metab. 2024:101954.
26. Torresan F, Crimi F, Ceccato F, Zavan F, Barbot M, Lacognata C et al. Radiomics: a new tool to differentiate adrenocortical adenoma from carcinoma. BJS Open. 2021;5(1).
27. . Elmohr MM, Fuentes D, Habra MA, Bhosale PR, Qayyum AA, Gates E, et al. Machine learning-based texture analysis for differentiation of large adrenal cortical tumours on CT. Clin Radiol. 2019;74(10):818. e1-e7.
28. Singh Y, Kelm ZS, Faghani S, Erickson D, Yalon T, Bancos I, et al. Deep learning approach for differentiating indeterminate adrenal masses using CT imaging. Abdom Radiol (NY). 2023;48(10):3189-94.
29. Crimi F, Quaia E, Cabrelle G, Zanon C, Pepe A, Regazzo D et al. Diagnostic accuracy of CT texture analysis in adrenal masses: a systematic review. Int J Mol Sci. 2022;23(2).
30. Zhang H, Lei H, Pang J. Diagnostic performance of radiomics in adrenal masses: a systematic review and meta-analysis. Front Oncol. 2022;12:975183.
31. Kong J, Zheng J, Wu J, Wu S, Cai J, Diao X, et al. Development of a radiomics model to diagnose pheochromocytoma preoperatively: a multicenter study with prospective validation. J Transl Med. 2022;20(1):31.
32. Kocak B, Baessler B, Bakas S, Cuocolo R, Fedorov A, Maier-Hein L, et al. Check- List for evaluation of radiomics research (CLEAR): a step-by-step reporting guideline for authors and reviewers endorsed by ESR and EuSoMI. Insights Imaging. 2023;14(1):75.
33. Kocak B, Akinci D’Antonoli T, Mercaldo N, Alberich-Bayarri A, Baessler B, Ambrosini I, et al. METhodological radiomics score (METRICS): a quality scoring tool for radiomics research endorsed by EuSoMII. Insights Imaging. 2024;15(1):8.
34. Zwanenburg A, Vallieres M, Abdalah MA, Aerts H, Andrearczyk V, Apte A, et al. The image biomarker standardization initiative: standardized quantita- tive radiomics for high-throughput image-based phenotyping. Radiology. 2020;295(2):328-38.
35. Orlhac F, Frouin F, Nioche C, Ayache N, Buvat I. Validation of a method to compensate multicenter effects affecting CT radiomics. Radiology. 2019;291(1):53-9.
36. Koo TK, Li MY. A guideline of selecting and reporting intraclass correlation coefficients for reliability research. J Chiropr Med. 2016;15(2):155-63.
37. Chen T, Guestrin C, XGBoost. Proceedings of the 22nd ACM SIGKDD Inter- national Conference on Knowledge Discovery and Data Mining. 2016. pp. 785-94.
38. Lundberg SM, Nair B, Vavilala MS, Horibe M, Eisses MJ, Adams T, et al. Explain- able machine-learning predictions for the prevention of hypoxaemia during surgery. Nat Biomed Eng. 2018;2(10):749-60.
39. Chinese Urological Doctor Association. Chinese expert consensus on diagno- sis and treatment of adrenocortical carcinoma. J Mod Urol. 2021;26(11):902- 8. (in Chinese).
40. Chinese Society of Endocrinology. Expert consensus on the diagnosis and treatment of pheochromocytoma and paraganglioma. (2020). 2020. (in Chinese).
41. Riley RD, Snell KIE, Archer L, Ensor J, Debray TPA, van Calster B, et al. Evaluation of clinical prediction models (part 3): calculating the sample size required for an external validation study. BMJ. 2024;384:e074821.
42. Mayo-Smith WW, Song JH, Boland GL, Francis IR, Israel GM, Mazzaglia PJ, et al. Management of incidental adrenal masses: a white paper of the ACR incidental findings committee. J Am Coll Radiol. 2017;14(8):1038-44.
43. Kline GA, Boyd J, Sadrzadeh HSM, Leung AA. Inpatient measurements of urine metanephrines are indistinguishable from pheochromocytoma: retro- spective cohort study. Am J Med. 2021;134(8):1039-e463.
44. Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, they are data. Radiology. 2016;278(2):563-77.
45. Cheng Y, Kou W, Zhu D, Yu X, Zhu Y. Future directions in diagnosis, prognosis and disease monitoring of adrenocortical carcinoma: novel non-invasive biomarkers. Front Endocrinol (Lausanne). 2021;12:811293.
46. Elmohr MM, Fuentes D, Habra MA, Bhosale PR, Qayyum AA, Gates E, et al. Machine learning-based texture analysis for differentiation of large adrenal cortical tumours on CT. Clin Radiol. 2019;74(10):e8181-7.
47. Rudin C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat Mach Intell. 2019;1(5):206-15.
48. Reyes M, Meier R, Pereira S, Silva CA, Dahlweid FM, von Tengg-Kobligk H, et al. On the interpretability of artificial intelligence in radiology: challenges and opportunities. Radiol Artif Intell. 2020;2(3):e190043.
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.