EJE
More than meets the eye: predicting adrenocortical carcinoma outcomes with pathomics
Jianqiu Kong, 1,2,3,1 Mingli Luo, 1,2,3,+ Yi Huang, 4,+ Ying Lin,5 Kaiwen Tan,6,7 Yitong Zou, 1,2,3 Juanjuan Yong,8 Sha Fu,8,9 Shaoling Zhang,5* Xinxiang Fan,1,2,3,* and Tianxin Lin 1,2,3,* [D
1Department of Urology, Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, Guangzhou 510120, Guangdong, PR China
2Guangdong Provincial Key Laboratory of Malignant Tumor Epigenetics and Gene Regulation, Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, Guangzhou 510120, Guangdong, PR China
3Guangdong Provincial Clinical Research Center for Urological Diseases, Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, Guangzhou 510120, Guangdong, PR China
4Department of Urology, The Third People’s Hospital of Chengdu, The Affiliated Hospital of Southwest Jiaotong University, Chengdu 610014, Sichuan, PR China
5Department of Endocrinology, Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, Guangzhou 510120, Guangdong, PR China
6Yunnan Key Laboratory of Artificial Intelligence, Kunming University of Science and Technology, Kunming 650500, China
7Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Yunnan, Kunming 650500, PR China 8Department of Pathology, Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, Guangzhou 510120, Guangdong, PR China
9Cellular and Molecular Diagnostics Center, Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, Guangzhou 510120, Guangdong, PR China
*Corresponding authors: Department of Urology, Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, 107 Yan Jiang West Road, Guangzhou 510120, Guangdong, PR China. Email: lintx@mail.sysu.edu.cn; Department of Endocrinology, Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, 107 Yan Jiang West Road, Guangzhou 510120, Guangdong, PR China. Email: zhshaol@mail.sysu.edu.cn; Department of Urology, Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, 107 Yan Jiang West Road, Guangzhou 510120, Guangdong, PR China. Email: fanxinx3@mail.sysu.edu.cn
Abstract
Background: Adrenocortical carcinoma (ACC) is a rare, aggressive malignancy with high recurrence rates and poor prognosis. Current prognostic models are inadequate, highlighting the need for innovative diagnostic tools. Pathomics, which utilizes computer algorithms to analyze whole- slide images, offers a promising approach to enhance prognostic models for ACC.
Methods: A retrospective cohort of 159 patients who underwent radical adrenalectomy between 2002 and 2019 was analyzed. Patients were divided into training (N= 111) and validation (N=48) cohorts. Pathomics features were extracted using an unsupervised segmentation method. A pathomics signature (PSAcc) was developed through LASSO-Cox regression, incorporating 5 specific pathomics features.
Results: The PSAcc showed a strong correlation with ACC prognosis. In the training cohort, the hazard ratio was 3.380 (95% CI, 1.687-6.772, P <. 001), and in the validation cohort, it was 3.904 (95% CI, 1.039-14.669, P <. 001). A comprehensive nomogram integrating PSAcc and M stage significantly outperformed the conventional clinicopathological model in prediction accuracy, with concordance indexes of 0.779 versus 0.668 in the training cohort (P =. 002) and 0.752 versus 0.603 in the validation cohort (P =. 003).
Conclusions: The development of a pathomics-based nomogram for ACC presents a superior prognostic tool, enhancing personalized clinical decision making. This study highlights the potential of pathomics in refining prognostic models for complex malignancies like ACC, with implications for improving prognosis prediction and guiding treatment strategies in clinical practice.
Keywords: adrenocortical carcinoma, pathomics, prognostic model, nomogram
Significance
This study demonstrates the potential of pathomics as a powerful tool for developing more accurate prognostic models in adrenocortical carcinoma (ACC). By combining pathological features with clinical variables, our model enhances the pre- cision of prognosis prediction, which can help guide personalized treatment strategies and improve patient outcomes. This research represents a significant advancement in the application of digital pathology and machine learning to complex ma- lignancies, offering a new approach to address the limitations of traditional prognostic methods in ACC.
+ These are the co-first authors.
Introduction
Adrenocortical carcinoma (ACC) is an infrequent and malig- nant neoplasm emanating from the adrenal cortex, with an an- nual incidence rate ranging between 0.5 and 2 cases per million population.1 It is important to highlight that ACC can manifest at any age and is characterized by its aggressive invasiveness. From a clinical perspective, ACC often presents itself as local invasion into the kidney and inferior vena cava or distant metastasis involving the liver and lungs.2
The critical therapeutic strategy for ACC is complete sur- gical resection, which stands as the sole curative approach for this condition.3 Nevertheless, the prognosis of ACC is frequently poor, with nearly 60% of patients experiencing recurrence after radical surgery, and the median overall sur- vival (OS) is reported to be ~3-4 years.4,5 Therefore, for pa- tients with ACC who have undergone radical surgical resection, it is essential to establish a reliable prognostic model, which can aid in developing personalized compre- hensive treatment plans, including appropriate adjuvant therapy and follow-up intervals, and can also alleviate the patient’s anxiety.
Despite efforts to improve prognostication, accurately pre- dicting post-resection prognosis in ACC remains a significant clinical challenge. The European Network for the Study of Adrenal Tumors (ENSAT) staging system, revised in 2015, serves as a cornerstone for prognostic stratification of patients with ACC.6,7 Histoprognostic factors such as mitotic count and Ki67 have been identified as crucial for prognosis in ACC. Studies have shown that the mitotic rate is a strong indicator of patient outcome, with tumors having more than 20 mitotic figures per 50 high-power fields classified as high grade and associated with poorer prognosis.8 Similarly, Ki67, a marker of cell proliferation, has been shown to have significant prognostic value, with higher Ki67 levels correlating with increased risk of recurrence and reduced OS.9
Nonetheless, given the great heterogeneity of ACC, the cur- rently available prognostic models still lack sufficient accuracy in predicting post-surgical survival.5 This highlights the urgent need for discovering new tools to provide more precise prog- nostic information for ACC. Experienced pathologists’ evalu- ation of hematoxylin-eosin (H&E)-stained slides according to the Weiss criteria is indispensable for the diagnosis of ACC, suggesting that the pathological tissue may contain a wealth of potential information reflecting the biological characteris- tics and prognostic features of ACC.
As digital pathology advances, a novel technology named “pathomics” has emerged in recent years.1º Pathomics utilizes computer algorithms to analyze high-resolution whole-slide images (WSIs) and extract quantitative image features in order to identify useful indicators to assist in clinical decision mak- ing.11-13 Compared with traditional methods of visual assess- ment and manual quantification, pathomics improves the efficiency and accuracy of information extraction of patho- logical image and uncovers a more diverse range of patho- logical features. The accuracy of predicting survival using pathomics has been demonstrated in fields such as gastric can- cer, renal cell carcinoma, and bladder cancer.14-16 However, there is currently no research exploring the prognostic value of pathomics in ACC.
In this study, we endeavored to extract useful patho- logical features from H&E-stained slides of ACC using
pathomics and to combine these features with relevant clin- ical variables to develop a more comprehensive and accurate prognostic model for ACC. We also assessed the value of this model in predicting postoperative OS in patients with ACC, enhancing our understanding of the role of pathomics in the prognostic evaluation of ACC. This work is an ex- tension of our previous efforts and contributes to refining the approach to personalized medicine in the management of ACC.
Methods
Participants and sample collection
This is a retrospective cohort study that recruited 159 patients with ACC who underwent radical adrenalectomy in the Sun Yat-Sen Memorial Hospital (N=80, diagnosed between December 2002 and May 2019) and The Cancer Genome Atlas (TCGA) (N = 79). The patients were then randomly div- ided into a training cohort (N = 111) and a validation cohort (N=48) in a 7:3 ratio. All patients met the following inclusion criteria: (1) histologically diagnosed with ACC based on Weiss criteria and additional pathological and clinical assessments and received radical surgical resection, (2) complete clinicopa- thological data and follow-up information of OS were available, (3) H&E-stained slides of tumor tissues with a thickness of 5 um were obtained, and (4) no history of other malignant tumors. The patients were followed up every 3 months during the first 2 years post-surgery, and every 6 months thereafter until the last follow-up date. The OS was defined as the time interval between surgery and death or the last follow-up date, and the survival status was recorded at the last follow-up. Baseline clinicopathological data were collected, including age, sex, TNM stage, and laterality of tu- mor. The 8th edition of the TNM staging system for adrenal tumors, introduced by the American Joint Committee on Cancer (AJCC) in 2017, was used.17
All tumor samples were collected by surgical resection and processed into formalin-fixed and paraffin-embedded sections for subsequent H&E staining. Slides were selected based on their ability to represent the key pathological features of adre- nocortical carcinoma (ACC), including regions with high tu- mor cell density, cellular atypia, and mitotic activity. Areas with extensive necrosis, hemorrhage, or artifacts were ex- cluded to ensure accurate pathomics feature extraction. All selections were made by experienced pathologists with over 10 years of experience in endocrine pathology, ensuring con- sistency and representativeness across all samples. These H&E-stained slides were then scanned at 40x magnification using the KF-PRO-120 scanner (KFBIO Technology for Health) to create WSIs, and the image files were saved as.svs format.
This study was conducted in accordance with the principles outlined in the Declaration of Helsinki and approved by the Institutional Review Committee of Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University. All patients who partici- pated in the study provided written informed consent prior to surgery. This consent included explicit permission for the collection, use, and sharing of their clinicopathological data for scientific research purposes. For the data obtained from TCGA, we utilized publicly available datasets. The use of TCGA data complies with the database’s policies on data ac- cess and publication, ensuring that patient confidentiality and ethical guidelines are maintained.
Figure 1. The workflow of pathomics processing and machine learning analysis in this study. WSI, whole-slide image.
ACC-Pathomics model
Pathomic
Medical Record
Area
Major axis length Minor axis length Ratio ..
.
Cell-level features
Age
Gender
Area
T stage N stage
Histogram
M stage Location
10-bin histogram
Distribution stastics
Tumor
Mean, SD, skewness, kurtosis, entropy
ENSAT stage Weiss score Mitotic count Hormone status
Adrenocortical carcinoma
Whole-slide imaging
Nucleus segmentation
Patient-level features (n = 150)
Modeling
Points
Manage
Não
2
-
-18
-12
-04
-48
-64
Ha
Total Points
6
.
10
14
1
S
1-year survival
3-year survival
00 OT 06 05 04 03 02 01015
5-year sundal
9.9
GÀ 0.7 06 06 04 03 02 01016
https://sysumh-urology.shinyapps.io/ACC_Pathomics_outcome/
Pipeline of pathomics feature extraction
Before feature extraction, it is necessary to segment each WSI into non-overlapping patches with the size of 2000 x 2000 pixels to facilitate subsequent processing.18 Then, the image feature extraction pipeline contains 3 steps: nucleus segmenta- tion, cell-level feature extraction, and aggregation of cell-level features into patient-level features. The workflow for extract- ing histopathological features from WSIs is shown in Figure 1.
Firstly, we utilized an unsupervised segmentation method proposed by Phoulady et al.19 to segment all nuclei in the WSIs. This method efficiently and accurately identifies the nu- clei in histopathological images without parameter learning or training data for adaptive parameters. Following that, 10 types of cell-level features were extracted for each segmented nucleus, which quantitatively reflected the size, texture, shape, and density characters of each nucleus. The specific features comprised the nucleus area, the lengths of the major and minor axes of the nucleus and their ratios, the mean pixel values of nucleus in the red, green, and blue channels (rMean, gMean, and bMean, respectively), and the mean, maximum, and min- imum distances (distMean, distMax, and distMin, respective- ly) to neighboring nuclei in the Delaunay triangulation graph.20
Subsequently, 10 histogram features were constructed for each cell-level feature. Specifically, each histogram consists of 10 bins, with each bin representing the frequency of a par- ticular value range within that cell-level feature category.20 For instance, the 10 histogram features that correspond to the features of nucleus area are named Area_bin1 through
Area_bin10. Area_bin1 represents the percentage of nuclei with very small areas in the WSI, while Area_bin10 represents the percentage of nuclei with very large areas. In addition, we computed 5 distribution statistical features for each cell-level feature, namely mean, SD, skewness, kurtosis, and entropy.20
Thus, the aforementioned approach extracted a total of 150 patient-level features for each WSI (10x 15). Detailed feature extraction procedures and related code are described in materials and elsewhere. 18,20
Construction and evaluation of pathomics signature
In the training cohort, the least absolute shrinkage and selec- tion operator (LASSO)-Cox regression model was employed to identify pathomics features related to ACC prognosis from 150 extracted features. This model, based on advanced machine learning algorithms, has been widely used for correl- ation analysis between omics features and survival.14 The LASSO-Cox regression analysis also matched the selected pathomics features with appropriate weighting coefficients, and the resulting pathomics signature (PSACc) was a linear model containing these features and coefficients, which could be presented as:
PSACC = a1F1 + a2F2 + … + a;Fi,
where Fi denotes the selected pathomics features associated with survival and aj represents the corresponding LASSO coef- ficient. The formula for calculating the PSACc in the validation cohort was consistent with that in the training cohort.
X-tile software (version 3.6.1, Yale University School of Medicine, New Haven, CT, USA) was used to calculate an op- timal cutoff value to classify the patients with ACC into high-PSACC and low-PSACc groups in the training cohort. X-tile employs a minimum P-value approach to determine the optimal cutoff point by assessing all potential divisions of the data and selecting the one that provides the most signifi- cant prognostic separation. The same cutoff value was applied to the validation cohort. Then, the Kaplan-Meier method and log-rank test were used to investigate the differences in OS be- tween the high- and low-PSACc groups. The receiver operating characteristic (ROC) curve and area under the curve (AUC) were used to evaluate the sensitivity and specificity of the asso- ciation between PSACC and OS in all 159 cases.
Development and validation of prognostic model
Univariate and multivariate Cox regression analyses were con- ducted in the training cohort to detect independent predictors for OS. For multivariable Cox regression, backward stepwise
selection was adopted with the Akaike information criterion (AIC) used to determine the variables to be included in the model. Based on these independent predictors, a novel nomo- gram was developed for predicting the likelihood of 1-, 3-, and 5-year OS (Figure 1). Then, Harrell’s concordance index (C-index) and ROC curve were used to assess the discrimin- ation, specificity, and sensitivity of the nomogram.21 Concurrently, calibration curves were used to measure the agreement level between the predicted and actual OS.22 The decision curve analysis (DCA) was performed to assess the clinical usefulness and to visually demonstrate the net prof- it for clinical decisions.23 Finally, internal validation was con- ducted in the validation cohort to further appraise the performance of the nomogram.
Statistical analysis
In this study, all standardized statistical analyses were conducted using R software (version 4.0.2, https://www. r-project.org/). Details of the R packages and their associated
| Variables | Training (N=111) | Validation (N=48) | Statistic | P |
|---|---|---|---|---|
| Age | 46 ±16 | 51 ± 14 | -2.032 | 0.045ª |
| Sex | .849 | .357 | ||
| Female | 63 (56.76%) | 31 (64.58%) | ||
| Male | 48 (43.24%) | 17 (35.42%) | ||
| T stage | 7.909 | .048 | ||
| T1 | 10 (9.01%) | 6 (12.5%) | ||
| T2 | 56 (50.45%) | 20 (41.67%) | ||
| T3 | 25 (22.52%) | 5 (10.42%) | ||
| T4 | 20 (18.02%) | 17 (35.42%) | ||
| N stage | 0.000 | .987 | ||
| N0 | 100 (90.09%) | 44 (91.67%) | ||
| N1 | 11 (9.91%) | 4 (8.33%) | ||
| M stage | 0.310 | .577 | ||
| M0 | 92 (82.88%) | 38 (79.17%) | ||
| M1 | 19 (17.12%) | 10 (20.83%) | ||
| Laterality | 0.207 | .650 | ||
| Left | 69 (62.16%) | 28 (58.33%) | ||
| Right | 42 (37.84%) | 20 (41.67%) | ||
| ENSAT stage | 5.574 | .125 | ||
| I | 9 (8.11%) | 6 (12.50%) | ||
| II | 50 (45.05%) | 19 (39.58%) | ||
| III | 25 (22.52%) | 5 (10.42%) | ||
| IV | 27 (24.32%) | 18 (37.50%) | ||
| Weiss score | 13.622 | .136 | ||
| 1 | 2 (1.80%) | 3 (6.25%) | ||
| 2 | 5 (4.50%) | 5 (10.42%) | ||
| 3 | 21 (18.92%) | 6 (12.50%) | ||
| 4 | 16 (14.41%) | 6 (12.50%) | ||
| 5 | 23 (20.72%) | 6 (12.50%) | ||
| 6 | 11 (9.91%) | 8 (16.67%) | ||
| 7 | 7 (6.31%) | 4 (8.33%) | ||
| 8 | 9 (8.11%) | 0 (0.00%) | ||
| 9 | 2 (1.80%) | 3 (6.25%) | ||
| Not available | 15 (13.51%) | 7 (14.58%) | ||
| Mitotic count | 19±23 | 36 ±39 | 60.375 | .008ª |
| Hormone secretion status | 1.812 | .981 | ||
| Androgen | 40 (36.04%) | 18 (37.50%) | ||
| Cortisol | 20 (18.02%) | 10 (20.83%) | ||
| Estrogen | 2 (1.80%) | 0 (0.00%) | ||
| Mineralocorticoids | 5 (4.50%) | 1 (2.08%) | ||
| None | 40 (36.04%) | 18 (37.50%) | ||
| Unknown | 4 (3.60%) | 1 (2.08%) |
P-values were derived from the univariate association analyses between the training and validation sets. ªP <. 05.
methods used in this study are provided in the materials. The differences between categorical variables were evaluated using x2 test or Fisher exact test. For continuous variables with non- normal distribution, the Mann-Whitney U test was employed. P-value <. 05 was considered as statistically significant.
Results
Baseline characteristics of all participants
The baseline characteristics of the patients with ACC are sum- marized in Table 1 for both the training (N = 111) and valid- ation (N=48) cohorts. In a total of 159 patients with ACC, ~60% were female and 40% were male, with a median (IQR) age of 49 (37-59) years. The majority of patients had clinical Stages I-III. The median (IQR) follow-up time for all patients was 1.12 (2.33-5.09) years. While there were slight differences between the training and validation cohorts in terms of age, T stage, and mitotic count, these variations were accounted for in the analysis. Overall, the baseline characteristics indicate that the 2 cohorts were sufficiently balanced for the purpose of model construction and validation.
Construction of PSAcc and its association with prognosis
In the training cohort, LASSO-Cox regression analysis was performed on the 150 pathomics features of all patients, and the minimum value of the tuning parameter lambda (2) was de- termined to be 0.073, as shown in Figure S1A and B. Thus, 5 pathomics features most associated with OS (distMax_ bin10, gMean_skewness, gMean_entropy, rMean_bin6, and rMean_entropy) were identified for constructing the PSACC, as shown in Figure S1C. The PSACc is calculated using the equation: PSACC= distMax_bin10x(-50.0011273) + gMean_skewness x (-0.3548547) + gMean_entropy x (-0.2598667)+rMean_bin6 x(-0.3806724) + rMean_entropy x (-0.1431038). By calculat- ing the Pearson correlation coefficients between these features, the independence of each selected feature was verified (Figure S2). Figure 2A presents the representative histo- pathological tiles for 5 pathomics features (disMax_bin10, gMean_skewness, gMean_entropy, rMean_bin6, and rMean_ entropy) with high and low values, demonstrating the visual dif- ferences in staining intensity and nuclear distribution. Specifically, distMax_bin10 reflects variations in nuclear distribution, with higher values indicating sparser nuclear
A
High
Low
B
distMax_bin10
%
”
gMean_skewness
.
.
1.00
0.75
gMean_entropy
True positive rate
.
0.50
0.25
AUC at 1 years = 0.786 (0.697 - 0.874)
·
AUC at 3 years = 0.654 (0.540 - 0.767)
rMean_bin6
0.00
AUC at 5 years = 0.640 (0.523 - 0.758)
0.00
0.25
0.50
0.75
1.00
False positive rate
9
rMean_entropy
Adrenocortical carcinoma
A
Training set
B
Validation set
-1.00
Risk Score
-0.5-
Risk Group
Risk Score
-1.0
-1.25
Risk Group
Low
-1.50
Low
-1.5
High
High
-1.75
Risk Group
-2.00
Risk Group
Low
Low
gMean_entropy
High
gMean_entropy
High
rMean_entropy
Expression
rMean_entropy
Expression
gMean_skewness
2
gMean_skewness
distMax_bin10
2
0
distMax_bin10
rMean_bin6
-2
rMean_bin6
0
PS
ACC
-4
PS
-2
ACC
status
status
C
Training set
D
Validation set
1.00
1.00
Survival probability
0.75
Survival probability
0.75
0.50
0.50
0.25
P = 0.00061
0.25
P = 0.0021
HR (95%) = 3.380 (1.687- 6.772)
HR (95%) = 3.904 (1.039- 14.669)
0.00
0.00
0
2.5
5
7.5
10
12.5
0
2
4
6
8
Time
Time
Number at risk
Number at risk
Low risk
58
32
21
11
3
0
Low risk
34
20
11
4
2
High risk
53
22
12
4
2
2
High risk
14
2
0
0
0
0
2.5
5
7.5
10
12.5
0
2
4
6
8
Time
Time
arrangements and lower values suggesting more tightly packed nuclei, which may correlate with increased proliferative capacity; gMean_entropy reflects the complexity and randomness in the distribution of staining intensities, highlighting the variability within the tissue structure. In contrast, gMean_skewness measures the asymmetry in these distributions, providing insights into the non-uniformity of chromatin texture. These parameters, while related to staining characteristics, offer distinct perspectives on the tissue morphology and underlying het- erogeneity. The rMean_bin6 reflects specific ranges of eosin stain- ing intensity, highlighting consistent staining patterns associated with cytoplasmic components or extracellular matrix. In contrast, rMean_entropy measures the variability in staining intensity, indi- cating broader heterogeneity within the tissue’s cytoplasmic and extracellular structures. The ROC curves of PSACC are presented in Figure 2B, and the AUCs (95% CI) for 1-, 3-, and 5-year OS of all patients are 0.786 (0.697-0.874), 0.654 (0.540-0.767), and 0.640 (0.523-0.758), respectively.
Classification of PSAcc and survival analysis
Through the utilization of X-tile software, an optimal cutoff value of -1.31 was determined for PSACC in the training
cohort (Figure S3), based on which the training cohort was then classified into high-PSACC (N=53) and low-PSACC (N =58) groups (Figure 3A). In the validation cohort, the identical cutoff value was used to classify 14 and 34 patients into the high- and low-PSACC groups, respectively (Figure 3B). The survival analysis of the training cohort indi- cated that patients with high PSACC had a hazard ratio (HR) of 3.380 (95% CI, 1.687-6.772, P <. 001) to those with low PSACC (Figure 3C). Consistent results were obtained in the validation cohort [HR (95% CI): 3.904 (1.039-14.669), P <. 001] (Figure 3D).
Development and validation of prognostic model
In order to identify independent prognostic factors, we con- ducted Cox regression analysis on potential clinical risk factors and PSACC in the training cohort (Table 2). The univariate ana- lysis demonstrated that the ENSAT stage, PSACC, and M stage exhibited significant correlation with OS (P <. 05). However, in the multivariate analysis, only PSACC and M stage remained statistically significant (P <. 05), indicating the independent predictive capacity of these 2 variables for OS. Moreover, the distribution of PSACC among patients with different M
| Variables | Univariate cox regression | Multivariate cox regression Pathomics model | Multivariate cox regression Clinical model | |||
|---|---|---|---|---|---|---|
| HR (95% CI) | P | HR (95% CI) | P | HR (95% CI) | P | |
| The pathomics signature (per 1 increase) | 9.555 (3.067-29.770) | <. 001ª | 5.235 (1.906-14.380) | .001ª | ||
| Age | 1.022 (1.000-1.044) | .053 | ||||
| Sex (Male vs. Female) | 0.907 (0.451-1.825) | .785 | ||||
| T stage | ||||||
| T1 | Reference | |||||
| T2 | 0.523 (0.144-1.904) | .326 | ||||
| T3 | 1.945 (0.499-7.579) | .338 | ||||
| T4 | 2.920 (0.814-10.473) | .100 | ||||
| N stage (N0 vs. N1) | 2.363 (0.971-5.750) | .058 | ||||
| M stage (M0 vs. M1) | 6.970 (3.314-14.660) | <. 001ª | 6.294 (2.905-13.640) | <. 001ª | 6.970 (3.314-14.660) | <. 001ª |
| Laterality (Left vs. Right) | 1.005 (0.500-2.022) | .988 | ||||
| ENSAT stage | ||||||
| I | Reference | |||||
| II | 0.769 (0.166 - 3.565) | .737 | ||||
| III | 1.525 (0.291 - 7.986) | .617 | ||||
| IV | 4.7306 (1.078 - 20.753) | .039 | ||||
| Weiss score | ||||||
| 1 | Reference | |||||
| 2 | 0.182 (0.011 - 2.978) | .232 | ||||
| 3 | 0.374 (0.044 - 3.155) | .366 | ||||
| 4 | 0.072 (0.004 - 1.159) | .064 | ||||
| 5 | 0.673 (0.085 - 5.354) | .709 | ||||
| 6 | 0.253 (0.023 - 2.811) | .264 | ||||
| 7 | 0.281 (0.025 - 3.130) | .302 | ||||
| 8 | 0.806 (0.094 - 6.926) | .844 | ||||
| 9 | 0.593 (0.037 - 9.517) | .712 | ||||
| Not Available | 0.358 (0.041 - 3.115) | .352 | ||||
| Mitotic count | 1.011 (0.998 - 1.025) | .106 | ||||
| Hormone secretion status | ||||||
| Androgen | Reference | |||||
| Cortisol | 0.636 (0.248 - 1.629) | .346 | ||||
| Estrogen | 0.000 (0.000 - Inf) | .998 | ||||
| Mineralocorticoids | 0.000 (0.000 - Inf) | .998 | ||||
| None | 0.546 (0.241 - 1.237) | .147 | ||||
| Unknown | 0.849 (0.195 - 3.700) | .827 | ||||
ªP <. 05.
Abbreviation: HR, hazard ratio.
stages was investigated, and no significant interaction was ob- served (Figure S4). Therefore, a pathomics nomogram incorp- orating PSACC and M stage was developed to predict the likelihood of 1-, 3-, and 5-year OS for patients with ACC (Figure 4A). Additionally, we developed an online tool based on this nomogram, which allows for easy access to survival prediction by scanning a QR code with a smartphone (Figure 4A).
ROC curves, calibration curves, DCA, and C-index were used to evaluate the performance of the pathomics nomogram. In the training cohort, the AUCs (95% CI) corresponding to 1-, 3-, and 5-year OS were 0.792 (0.668-0.916), 0.795 (0.676-0.914), and 0.836 (0.730-0.943), respectively (Figure 4B). Similarly, the AUCs (95% CI) for 1-, 3-, and 5-year OS in the validation cohort were 0.796 (0.643-0.949), 0.759 (0.571-0.948), and 0.837 (0.645-1.021), respectively (Figure 4C). The calibration curves illustrated that the 1-, 3-, and 5-year OS predicted by the pathomics nomogram were in good agreement with the actual OS in both the training and validation cohorts (Figure S5), which demonstrated the precision of the nomogram in survival prediction. The results of DCA indicated that the use of pathomics nomogram for
clinical decision making could yield greater net benefits com- pared with the treat-all scheme or the treat-none scheme (Figure S6), highlighting the clinical applicability of this nomogram.
Comparison of pathomics nomogram and clinicopathological model
The distribution of various clinical and pathological variables, including survival status, is presented in Figure S7. Sankey dia- grams (Figure S7A-K) illustrate the relationship between dif- ferent clinical and pathological variables (such as hormone secretion status, Weiss score, and ENSAT stage) and survival status (Alive vs. Dead) in patients with ACC. These diagrams provide a comprehensive view of how these variables correlate with patient outcomes, further underscoring the prognostic value of the identified pathomics features.
To better demonstrate the superiority of the pathomics nomogram, we constructed an additional prediction model comprising only clinicopathological factors for comparison. According to the results of multivariate Cox regression, M stage was identified and included as a predictor in the
A
Points
0
1
2
3
4
5
6
7
8
9
10
M stage
M1
MO
PS
ACC
-2
-1.8
-1.6
-1.4
-1.2
-1
-0.8
-0.6
-0.4
-0.2
Total points
0
2
4
6
8
10
12
14
16
18
1-year survival
0.95
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
3-year survival
0.95
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1 0.05
5-year survival
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.05
B
C
1.00
1.00
0.75
0.75
True positive rate
True positive rate
0.50
0.50
0.25
0.25
AUC at 1 years = 0.792 (0.668 - 0.916)
AUC at 1 years = 0.796 (0.643 - 0.949)
AUC at 3 years = 0.795 (0.676 -0.914)
AUC at 3 years = 0.759 (0.571 - 0.948)
0.00
AUC at 5 years = 0.836 (0.730 - 0.943)
0.00
AUC at 5 years = 0.837 (0.654 - 1.021)
0.00
0.25
0.50
0.75
1.00
0.00
0.25
0.50
0.75
1.00
False positive rate
False positive rate
construction of the clinicopathological model. In the training cohort, the pathomics nomogram yielded a C-index (95% CI) of 0.779 (0.703-0.855), while the clinicopathological model yielded a C-index (95% CI) of 0.668 (0.586-0.750). The difference between the two holds statistical significance (P =. 002), indicating that the pathomics nomogram had su- perior prediction accuracy (Table S1). This conclusion found support in the results of the validation cohort, with C-indices of 0.752 and 0.603 for the pathomics nomogram and clinicopathological model, respectively.
Discussion
This is currently the first study to investigate the use of patho- mics techniques in extracting high-throughput features from H&E-stained slides for developing a prognostic model for ACC. In this study, we proposed an innovative pathomics sig- nature (PSACC) to assist clinicians in risk-stratifying patients with ACC. By further integrating clinical information with PSACC, we put forward a novel nomogram for predicting sur- vival in patients with ACC and substantiated via a series of rigorous validations that this pioneering tool holds great promise for enhancing prognostic accuracy in the aforemen- tioned patient population.
Owing to the low incidence and heterogeneity of ACC, there is still a dearth of high-level clinical research evidence to ascertain the most effective prognostic factors. So far, re- searchers have suggested several prognostic models on the ba- sis of differential combinations of clinical/histopathological information.24-27 Among these models, the 2015 revision of the ENSAT staging system is recommended by European guidelines as the most discriminative prognostic stratification tool for ACC.5 In a further multicenter study, the S-GRAS score, which combined ENSAT stage, Ki-67, tumor resection status, and other clinical data, demonstrated superior prog- nostic value and could forecast the efficacy of mitotane treat- ment.7 It can be noted from these prognostic models relying on conventional prognostic factors that: (1) The selection of prognostic factors remains a subject of controversy; for in- stance, multiple studies have demonstrated that age and tumor size function as independent prognostic factors for survival after radical resection of ACC,24-26 while certain research sug- gested that these variables are not prognostic.28,29 (2) The ac- curacy of most such models remains limited; even the aforementioned S-GRAS score, constructed on the basis of a large cohort population, exhibited a C-index below 0.8.7 (3) Whether the applicability of such models can be generalized to all individuals remains validation through prospective
studies. In recent years, markers derived from gene expression analysis and targeted molecular analysis have also shown good ability to prognosticate the outcome of ACC. The inte- gration of specific somatic mutations, Wnt/ß-catenin and p53 pathways alterations, and the promoter methylation in- dex into the S-GRAS score reportedly yielded exceedingly en- couraging predictive effects (AUC=0.872).30 Other studies have also indicated the tremendous potential of detecting spe- cific multigene panels in enhancing the accuracy of survival predictions for ACC.31 However, these cutting-edge technolo- gies related to genetic data analysis are cumbersome and cost- ly, hereby limiting their utilization and popularized in routine clinical practice. To overcome the limitations of existing prog- nostic methods, it is necessary to explore novel indicators that are more sensitive and universally applicable. At present, there is a lack of research regarding the prognostic information provided by pathomics techniques in the field of ACC. In our study, we found that the PSACC derived from H&E-stained slides contributes to prognostic stratification in ACC. The OS between high- and low-PSACc groups showed a significant difference. Compared with prognostic tools pre- viously reported, our study introduced an innovative ap- proach to capture potential prognostic information at the micro-level by constructing PSACC based on the features of microscopic pathological images. It possessed the advantages of easy specimen collection and practicality. Furthermore, the composite prognostic model combining both PSACC and M stage demonstrated remarkable superiority in prediction accuracy (AUCs at 1-, 3- and 5-year were 0.792, 0.795, and 0.836, respectively). In future research, it is valuable to ex- plore the integration of PSACC with additional potential prog- nostic factors to construct more precise and personalized prognostic tools.
In clinical practice, conventional microscopic observation not only provides critical diagnostic information, but also of- fers valuable prognostic insights. For instance, the Weiss score, initially designed to differentiate adrenocortical aden- omas from ACC, has prognostic implications as it was estab- lished by comparing metastatic and non-metastatic tumors. 32 The Weiss score, which includes the mitotic index, and Ki67, an independent prognostic marker, are both well-established markers with significant prognostic value in ACC.9,33-35 Additionally, the Helsinki score, combining these factors, is also correlated with survival, further emphasizing the import- ance of integrating histopathological data in prognostic models.36,37 Nevertheless, this predicament has undergone a pivotal transformation in recent years. With the advent of digital pathology, the ability to scan and analyze WSIs has significantly enhanced our capacity to extract comprehensive pathomic features. This digitization process has been pivotal in the development of our nomogram, as it allows for the de- tailed analysis of tissue characteristics that were previously difficult to quantify.10,12,13 At present, pathomics provides a novel approach for predicting tumor survival.38,39 Chen et al.15 reported that pathomics features related to nuclear size and staining depth possess immense potential in enhan- cing survival prediction accuracy in clear cell renal cell carcin- oma. In the field of gastric cancer, pathomics features have also been established as indicators for postoperative risk stratification and as tools for predicting the benefits of adju- vant chemotherapy.14 Similarly, we developed and validated a pathomics-based PSACC for predicting survival of patients with ACC. The results showed that PSACC showed
satisfactory performance in distinguishing the high and low survival risks among patients with ACC, suggesting its poten- tial as an independent prognostic factor for ACC. Our study fills the gap in the research concerning the application of pathomics techniques to predict survival in ACC and demon- strates that pathomics can serve as an efficacious method sur- passing conventional pathological assessment to provide supplementary prognostic information in the treatment of ACC.
To our knowledge, no consensus has been reached to unify or standardize the extraction methods of pathomics fea- tures.12 In this study, an automated, unsupervised algorithm was employed to extract 150 quantitative image features from pathological images of ACC. For one thing, due to the relatively consistent composition of normal adrenal tissue in each pathological section, it is reasonable to perform feature extraction directly on the entire slide, without the need for manual annotation of the tumor region on WSIs. For another, extracting image features from the entire slide not only maxi- mizes coverage of the tumor region but also includes para- cancerous tissues and scattered tumor cells that may be missed during manual annotation. Therefore, compared with hand- crafted feature-based approaches,12 the unsupervised method we used can substantially save processing time and mitigate the influence of intratumoral heterogeneity on feature extrac- tion. In this study, the 150 features we extracted objectively re- flect a broad range of nuclear morphology in pathological sections, including size, texture, shape, and density. And the association of these well-defined features with ACC is highly interpretable. Specifically, through LASSO-Cox regression, the final set of pathomics features selected for the construc- tion of PSACC comprises distMax_bin10, gMean_skewness, gMean_entropy, rMean_bin6, and rMean_entropy. These features illustrate that within the wealth of information pro- vided by H&E-stained slides, the prognosis of ACC is mainly related to the density, staining depth and texture of the nuclei, which can be linked to cell divisions. Notably, among these features, distMax_bin10 contributes the most to the predic- tion of prognosis. Higher values of distMax_bin10 indicate a sparser nuclear distribution, while lower values suggest more densely packed nuclei, which are associated with in- creased cell proliferation. This finding is consistent with the negative coefficient of distMax_bin10 in our model, further validating the scientific rationale of our pathomic scoring sys- tem. The feature reflects variations in nuclear arrangement, and its significance lies in the observation that tumors with more densely packed nuclei tend to exhibit faster proliferative capacity, correlating with poorer prognosis. As such, distMax_bin10 serves as an important marker of nuclear dis- tribution and its relationship with tumor cell behavior in ACC, reinforcing the prognostic value of our model. The distMax_bin10 parameter may also reflect variations in cell density and size, including the presence of smaller compact eo- sinophilic cells and large oncocytic cells. While these cellular characteristics can influence tumor behavior, it is important to highlight that oncocytic ACC is generally associated with a better prognosis compared with conventional ACC. Oncocytic ACCs, which have a different prognostic profile compared with conventional ACCs, could still influence this parameter due to their distinct cellular morphology.37 The other 4 pathomics features (gMean_skewness, gMean_ entropy, rMean_bin6, and rMean_entropy) indicate that a higher proportion of deeply and unevenly stained nuclei
(possibly due to uneven chromatin distribution) correlates with a worse prognosis for ACC. This informative finding could be linked to mitosis or nucleolated nuclei, which are classical Weiss score parameters. The pathology images in- cluded in our study provide qualitative support for the identi- fied pathomic features. For instance, the images with high gMean_skewness values display noticeable chromatin hetero- geneity, which is likely related to nucleolar presence and activ- ity. These changes are indicative of increased tumor cell proliferation and malignant potential, both of which contrib- ute to more aggressive tumor behavior.4º Similarly, images il- lustrating high distMax_bin10 values show sparse nuclear distribution, reflecting reduced cell density and potential var- iations in tumor proliferation. These visual observations re- inforce the significance of these features in our prognostic model. These detailed descriptions highlight their relevance for pathological reports. The interpretability of these findings suggests that the prognostic information provided by PSACC may be based on classical pathological theories. For example, gMean_skewness may reflect nuclear atypia and chromatin clumping, rMean_bin6 could correspond to cytoplasmic density or stromal changes, and rMean_entropy may indicate tumor heterogeneity-all of which are classical histopatho- logical characteristics observed in malignancies.41
The TNM staging system is widely used in clinical practice to evaluate the progression and prognosis of a variety of ma- lignant tumors. Due to the relatively low incidence, it was not until 2003 that the AJCC established the TNM staging system for ACC.42 Despite evident limitations and the inabil- ity to provide a satisfactory prognostic stratification, 42,43 the TNM classification remains of significant reference value in the current assessment of ACC’s prognosis. Within this staging system, the presence of metastasis (M1) represents the strongest indicator of poor prognosis.5 In this study, we in- corporated TNM staging, along with gender, age, laterality, ENSAT staging, Weiss score, mitotic count, and cortisol secre- tion, into the correlation analysis with survival. Following uni- variate and multivariate Cox regression, the M stage was ultimately confirmed as the sole indicator significantly associ- ated with survival, which harmoniously aligns with the bio- logical characteristics of ACC. The C-index was 0.668 in the training cohort and 0.603 in the validation cohort when the prognostic model was developed based on M stage independ- ently. However, by integrating PSACC with M stage to create a pathomics nomogram, the corresponding C-index demon- strated a remarkable increase of 16.6% and 24.7%. Such in- crements further substantiate the notion that PSACC has the capacity to provide more comprehensive prognostic informa- tion, thereby significantly improving the accuracy of the pre- dictive model.
Although the reliability of the pathomics nomogram was demonstrated in different cohorts, there are certain limitations in this study. Firstly, this is a retrospective study, and thus, there are inevitably potential inherent biases and confounding factors. For example, the indirect impact of postoperative ad- juvant therapy (eg, receiving mitotane) on survival cannot be excluded. Secondly, all the patients included in this study were adults. Therefore, whether the same pathomics features can be used to forecast prognosis in pediatric patients with ACC remains to be validated. In the future, it is necessary to conduct prospective studies in a broader age range of individ- uals to further evaluate the clinical utility of PSACC, and it is worth considering the integration of PSACC with more
potential prognostic factors to construct a more comprehen- sive prognostic model.
In conclusion, our study successfully constructed PSACC using pathomics techniques and confirmed its significant correlation with the prognosis of ACC. Through the integration of PSACC with M stage, we developed and validated a nomogram, which shows commendable predictive accuracy and clinical utility and can serve as a user-friendly tool to guide clinicians in making personalized decisions for patients with ACC.
Supplementary material
Supplementary material is available at European Journal of Endocrinology online.
Funding
This work was supported by the National Natural Science Foundation of China (grant numbers 82203720, 82203188, 82002682, 81972731, 81773026, 81972383), the Guangzhou Municipal Basic Research Program Jointly Funded by City, University, and Enterprise Special Project (grant number 2024A03J0907), and the Natural Science Foundation of Guangdong Province (grant number 2024A1515013201).
Conflict of interest: None declared
Authors’ contributions
Jianqiu Kong (Conceptualization [equal], Formal analysis [equal], Funding acquisition [equal], Project administration [equal], Resources [equal]), Mingli Luo (Data curation [equal], Formal analysis [equal], Visualization [equal], Writing-original draft [equal]), Yi Huang (Data curation [equal], Investigation [equal], Visualization [equal], Writing -review and editing [equal]), Ying Lin (Data curation [equal], Formal analysis [equal], Validation [equal]), Kaiwen Tan (Formal analysis [equal], Methodology [equal], Software [equal]), Yitong Zou (Data curation [equal], Investigation [equal], Validation [equal]), Juanjuan Yong (Formal analysis [equal], Investigation [equal], Methodology [equal], Validation [equal]), Sha Fu (Formal analysis [equal], Investigation [equal], Software [equal], Validation [equal]), Shao-Ling Zhang (Data curation [equal], Validation [equal], Writing-review and editing [equal]), Xinxiang Fan (Data curation [equal], Methodology [equal], Supervision [equal], Validation [equal], Writing-review and editing [equal]), and Tianxin Lin (Conceptualization [equal], Funding acquisi- tion [equal], Project administration [equal], Writing-review and editing [equal]).
Data availability
The images and clinical data of the training cohort and valid- ation cohort will be made available at publication upon re- quest to the corresponding author.
References
1. Kerkhofs TM, Verhoeven RH, Van der Zwan JM, et al. Adrenocortical carcinoma: a population-based study on incidence and survival in the Netherlands since 1993. Eur J Cancer. 2013;49(11):2579-2586. https://doi.org/10.1016/j.ejca.2013.02.034
2. Else T, Kim AC, Sabolch A, et al. Adrenocortical carcinoma. Endocr Rev. 2014;35(2):282-326. https://doi.org/10.1210/er. 2013-1029
3. Datta J, Roses RE. Surgical management of adrenocortical carcinoma: an evidence-based approach. Surg Oncol Clin N Am. 2016;25(1):153-170. https://doi.org/10.1016/j.soc.2015.08. 011
4. Calabrese A, Basile V, Puglisi S, et al. Adjuvant mitotane therapy is beneficial in non-metastatic adrenocortical carcinoma at high risk of recurrence. Eur J Endocrinol. 2019;180(6):387-396. https:// doi.org/10.1530/EJE-18-0923
5. Fassnacht M, Dekkers OM, Else T, 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-G46. https://doi.org/10.1530/EJE- 18-0608
6. Libé R, Borget I, Ronchi CL, et al. Prognostic factors in stage III-IV adrenocortical carcinomas (ACC): an European Network for the Study of Adrenal Tumor (ENSAT) study. Ann Oncol. 2015;26(10):2119-2125. https://doi.org/10.1093/annonc/mdv329
7. Elhassan YS, Altieri B, Berhane S, et al. S-GRAS score for prognostic classification of adrenocortical carcinoma: an international, multi- center ENSAT study. Eur J Endocrinol. 2021;186(1):25-36. https://doi.org/10.1530/EJE-21-0510
8. Giordano TJ. The argument for mitotic rate-based grading for the prognostication of adrenocortical carcinoma. Am J Surg Pathol. 2011;35(4):471-473. https://doi.org/10.1097/PAS.0b013e31820b cf21
9. Beuschlein F, Weigel J, Saeger W, et al. Major prognostic role of Ki67 in localized adrenocortical carcinoma after complete resec- tion. J Clin Endocrinol Metab. 2015;100(3):841-849. https://doi. org/10.1210/jc.2014-3182
10. Gupta R, Kurc T, Sharma A, Almeida JS, Saltz J. The emergence of pathomics. Curr Pathobiol Rep. 2019;7(3):73-84. https://doi. org/10.1007/s40139-019-00200-x
11. Cao R, Yang F, Ma SC, et al. Development and interpretation of a pathomics-based model for the prediction of microsatellite instabil- ity in colorectal cancer. Theranostics. 2020;10(24):11080-11091. https://doi.org/10.7150/thno.49864
12. Bera K, Schalper KA, Rimm DL, Velcheti V, Madabhushi A. Artificial intelligence in digital pathology-new tools for diagnosis and precision oncology. Nat Rev Clin Oncol. 2019;16(11):703-715. https://doi.org/10.1038/s41571-019-0252-y
13. Niazi MKK, Parwani AV, Gurcan MN. Digital pathology and arti- ficial intelligence. Lancet Oncol. 2019;20(5):e253-e261. https://doi. org/10.1016/S1470-2045(19)30154-8
14. Chen D, Fu M, Chi L, et al. Prognostic and predictive value of a pathomics signature in gastric cancer. Nat Commun. 2022;13(1): 6903. https://doi.org/10.1038/s41467-022-34703-w
15. Chen S, Jiang L, Gao F, et al. Machine learning-based pathomics signature could act as a novel prognostic marker for patients with clear cell renal cell carcinoma. Br J Cancer. 2022;126(5):771-777. https://doi.org/10.1038/s41416-021-01640-2
16. Chen S, Jiang L, Zhang E, et al. A novel nomogram based on machine learning-pathomics signature and neutrophil to lymphocyte ratio for survival prediction of bladder cancer patients. Front Oncol. 2021 ;11:703033. https://doi.org/10.3389/fonc.2021.703033
17. Lloyd R, Osamura RY, Klöppel G & Rosai J. WHO classification of tumours of endocrine organs. 2017.
18. Cheng J, Han Z, Mehra R, et al. Computational analysis of patho- logical images enables a better diagnosis of TFE3 Xp11.2 transloca- tion renal cell carcinoma. Nat Commun. 2020;11(1):1778. https:/ doi.org/10.1038/s41467-020-15671-5
19. Phoulady HA, Goldgof DB, Hall LO, Mouton PR. Nucleus segmen- tation in histology images with hierarchical multilevel thresholding. In: Conference on Medical Imaging-Digital Pathology. Vol 9791. San Diego, CA: Spie-Int Soc Optical Engineering; 2016: 979111-1-919111-6.
20. Cheng J, Mao Y, Hong W, et al. Multimodal data analysis reveals that pancreatobiliary-type ampullary adenocarcinoma resembles pancreatic adenocarcinoma and differs from cholangiocarcinoma.
J Transl Med. 2022;20(1):272. https://doi.org/10.1186/s12967- 022-03473-w
21. Gerds TA, Kattan MW, Schumacher M & Yu C. Estimating a time- dependent concordance index for survival prediction models with covariate dependent censoring. Stat Med 2013:32(13):2173-2184. https://doi.org/10.1002/sim.5681
22. Kattan MW. Judging new markers by their ability to improve pre- dictive accuracy. J Natl Cancer Inst. 2003;95(9):634-635. https:// doi.org/10.1093/jnci/95.9.634
23. Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Making. 2006;26(6): 565-574. https://doi.org/10.1177/0272989X06295361
24. Kong J, Zheng J, Cai J, et al. A nomogram for individualized estima- tion of survival among adult patients with adrenocortical carcin- oma after surgery: a retrospective analysis and multicenter validation study. Cancer Commun (Lond). 2019;39(1):80. https:/ doi.org/10.1186/s40880-019-0426-0
25. Kim Y, Margonis GA, Prescott JD, et al. Nomograms to predict recurrence-free and overall survival after curative resection of adreno- cortical carcinoma. JAMA Surg. 2016;151(4):3065-3373. https://doi. org/10.1001/jamasurg.2015.4516
26. Li Y, Bian X, Ouyang J, Wei S, He M, Luo Z. Nomograms to predict overall survival and cancer-specific survival in patients with adrenocortical carcinoma. Cancer Manag Res. 2018;10: 6949-6959. https://doi.org/10.2147/CMAR.S187169
27. Ettaieb MHT, van Kuijk SMJ, de Wit-Pastoors A, et al. Development and internal validation of a multivariable prediction model for adrenocortical-carcinoma-specific mortality. Cancers (Basel). 2020;12(9):2720. https://doi.org/10.3390/cancers12092720
28. Canter DJ, Mallin K, Uzzo RG, et al. Association of tumor size with metastatic potential and survival in patients with adrenocortical carcinoma: an analysis of the national cancer database. Can J Urol. 2013;20:6915-6921.
29. Stojadinovic A, Ghossein RA, Hoos A, et al. Adrenocortical carcin- oma: clinical, morphologic, and molecular characterization. J Clin Oncol. 2002;20(4):941-950. https://doi.org/10.1200/JCO.2002. 20.4.941
30. Lippert J, Appenzeller S, Liang R, et al. Targeted molecular analysis in adrenocortical carcinomas: a strategy toward improved personal- ized prognostication. J Clin Endocrinol Metab. 2018;103(12): 4511-4523. https://doi.org/10.1210/jc.2018-01348
31. Lin X, Gu Y, Su Y, et al. Prediction of adrenocortical carcinoma re- lapse and prognosis with a set of novel multigene panels. Cancers (Basel). 2022:15:14. https://doi.org/10.3390/cancers15010014
32. Weiss LM. Comparative histologic study of 43 metastasizing and nonmetastasizing adrenocortical tumors. Am J Surg Pathol. 1984;8(3):163-169. https://doi.org/10.1097/00000478-198403000- 00001
33. Harrison LE, Gaudin PB, Brennan MF. Pathologic features of prog- nostic significance for adrenocortical carcinoma after curative re- section. Arch Surg. 1999;134(2):181-185. https://doi.org/10.1001/ archsurg. 134.2.181
34. Assié G, Antoni G, Tissier F, et al. Prognostic parameters of meta- static adrenocortical carcinoma. J Clin Endocrinol Metab. 2007;92(1):148-154. https://doi.org/10.1210/jc.2006-0706
35. Duregon E, Molinaro L, Volante M, et al. Comparative diagnostic and prognostic performances of the hematoxylin-eosin and phospho-histone H3 mitotic count and Ki-67 index in adrenocort- ical carcinoma. Mod Pathol. 2014;27(9):1246-1254. https://doi. org/10.1038/modpathol.2013.230
36. Pennanen M, Heiskanen I, Sane T, et al. Helsinki score-a novel model for prediction of metastases in adrenocortical carcinomas. Hum Pathol. 2015;46(3):404-410. https://doi.org/10.1016/j. humpath.2014.11.015
37. Mete O, Erickson LA, Juhlin CC, et al. Overview of the 2022 WHO classification of adrenal cortical tumors. Endocr Pathol. 2022; 33(1):155-196. https://doi.org/10.1007/s12022-022-09710-8
38. Klimov S, Miligy IM, Gertych A, et al. A whole slide image-based machine learning approach to predict ductal carcinoma in situ
(DCIS) recurrence risk. Breast Cancer Res. 2019;21(1):83. https:/ doi.org/10.1186/s13058-019-1165-5
39. Yu KH, Zhang C, Berry GJ, et al. Predicting non-small cell lung cancer prognosis by fully automated microscopic pathology image features. Nat Commun. 2016;7(1):12474. https://doi.org/10.1038/ncomms12474
40. Montanaro L, Treré D, Derenzini M. Nucleolus, ribosomes, and cancer. Am J Pathol. 2008;173(2):301-310. https://doi.org/10. 2353/ajpath.2008.070752
41. Zink D, Fischer AH, Nickerson JA. Nuclear structure in cancer cells. Nat Rev Cancer. 2004;4(9):677-687. https://doi.org/10. 1038/nrc1430
42. Fassnacht M, Johanssen S, Quinkler M, et al. Limited prognostic value of the 2004 International Union Against Cancer staging classification for adrenocortical carcinoma: proposal for a revised TNM classification. Cancer. 2009;115(2):243-250. https://doi. org/10.1002/cncr.24030
43. Lughezzani G, Sun M, Perrotte P, et al. The European Network for the Study of Adrenal Tumors staging system is prognostically super- ior to the international union against cancer-staging system: a North American validation. Eur J Cancer. 2010;46(4):713-719. https://doi.org/10.1016/j.ejca.2009.12.007