Check for updates
Contrast CT radiomic features add value to prediction of prognosis in adrenal cortical carcinoma
Jiacheng Liu1 . Wenhao Lin1 . Ling Yan2 . Jialing Xie3 . Jun Dai1 . Danfeng Xu1 . Juping Zhao1
Received: 26 June 2023 / Accepted: 9 October 2023 / Published online: 15 November 2023 @ The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2023
Abstract
Objective Adrenocortical carcinoma (ACC) is a rare and aggressive malignancy with poor prognosis due to high post- operative recurrence rates. The aim of this study is to develop a contrast CT radiomic feature-based prognosis prediction model for ACC and evaluate its performance by comparison with ENSAT staging system and S-GRAS score. Methods Included in this study were 39 ACC patients, from which we extracted 1411 radiomic features. Using cross- validated least absolute shrinkage and selection operator regression (cv-LASSO regression), we generated a radiomic index. Additionally, we further validated the radiomic index using both univariate and multivariate Cox regression analyses. We constructed a radiomic nomogram that incorporated the radiomic signature and compared it with ENSAT stage and S-GRAS score in terms of calibration, discrimination and clinical usefulnes.
Results In this study, the average progression free survival (PFS) of 39 patients was 20.4 (IQR 9.1-60.1) months and the average overall survival (OS) was 57.8 (IQR 32.4-NA). The generated radiomic features were significantly associated with PFS, OS, independent of clinical-pathologic risk factors (HR 0.16, 95%CI 0.02-0.99, p = 0.05; HR 0.20, 95%CI 0.04-1.07, p = 0.06, respectively). The radiomic index, ENSAT stage, resection status, and Ki67% index incorporated nomogram exhibited better performance for both PFS and OS prediction as compared with the S-GRAS and ENSAT nomogram (C-index: 0.75 vs. C-index: 0.68, p =0.030 and 0.67, p = 0.025; C-index: 0.78 vs. C-index: 0.72, p = 0.003 and 0.73, p = 0.006). Calibration curve analysis showed that the radiomics-based model performs best in predicting the two-year PFS and the three-year OS. Decision curve analysis demonstrated that the radiomic index nomogram outperformed the S-GRAS and ENSAT nomogram in predicting the two-year PFS and the three-year OS.
Conclusion The contrast CT radiomic-based nomogram performed better than S-GRAS or ENSAT in predicting PFS and OS in ACC patients.
Keywords Radiomic features . Adrenal cortical carcinoma · Prediction · Progression-free survival
These authors contributed equally: Jiacheng Liu, Wenhao Lin, Ling Yan
☒ Jun Dai dj11338@rjh.com.cn
☒ Danfeng Xu xdf12036@rjh.com.cn
☒ Juping Zhao zjp11317@rjh.com.cn
1 Department of Urology, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China
2 Department of Radiology, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China
3 Department of Pathology, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China
Highlights
· Adrenocortical carcinoma (ACC) is a rare and aggressive malignancy with poor prognosis. Current prognosis prediction models are limited in value.
· Contrast CT based radiomic features add value to ACC prognosis prediction.
· Nomogram incorporated with radiomic feature-based index outperformed the S-GRAS and ENSAT nomogram in term of prognosis prediction.
Introduction
Adrenocortical carcinoma (ACC) is an exceedingly rare and highly aggressive malignancy, with an annual incidence ranging from 1 to 2 cases per million individuals [1]. The 5-year overall survival (OS) rate varies widely, spanning from 15% to 84% depending on the disease stage [2, 3]. After the initial surgical resection, about 70-80% patients experienced recurrence [2, 4]. Nevertheless, there exists considerable heterogeneity in terms of progression-free survival (PFS) and 5-year OS. For instance, the estimated 5-year survival rate of the International Union Against Cancer (UICC) stage I patients is about 82% vs. 13% for stage IV [2, 5].
An accurate prognostic prediction would facilitate the enhanced management of postoperative patients by urolo- gists and oncologists. Previous studies have found that tumor/hormone-related symptoms, pre-operation TNM stage, resection status and pathological outcomes may all affect the PFS of post-surgical patients [1, 6-9]. In an effort to enhance the accuracy of prognosis prediction, several indices have been developed. The European Network for the Study of Adrenal Tumors (ENSAT) staging system is the currently most widely used tool for ACC prognosis prediction [5, 10]. However, the prediction ability of this staging system is limited by its inability to account for post- surgical factors, thus emasculating its comprehensive prognostic assessment ability. The S-GRAS score is a novel prognostic model developed from a multicenter cohort. By incorporating multiple factors, the S-GRAS score has demonstrated a promising accuracy in predicting the prog- nosis of ACC patients [7, 11].
However, the critical factor in these models, pre-surgical tumor state, is still assessed by a manual TNM stage. Radiomics is a novel field that focuses on the extraction and analysis of a vast array of quantitative imaging naked-eye invisible features from medical images such CT, MRI and PET. By leveraging large datasets of imaging features, radiomics enables the identification of correlations between specific radiomic signatures and various diagnostic and prognostic factors. The application of radiomics has demonstrated remarkable potential in the diagnosis and prognosis of various types of cancer [12-14]. In this study, we aim to develop an ACC prognosis prediction model by extracting and analyzing imaging features from contrast-
enhanced CT scans and evaluate its performance by com- paring it with ENSAT and S-GRAS in terms of the pre- diction ability in an independent cohort of ACC patients from a single center in China.
Materials and methods
This retrospective study was approved by the ethics com- mittee of Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, China. All experimental methods were performed in accordance with guidelines and regulations of the Ethics Committee of our center. Informed consent was obtained from all patients and/or legal guardians.
Patients
Initially recruited in this study were 187 patients who were pathologically diagnosed with ACC in Ruijin Hospital (Shanghai, China) between September 2010 and August 2022. Patients without pre-operative raw contrast CT ima- ges or clinical data, unsuitable for surgery, had previously experienced local or systematic therapy before surgery, missing information of survival outcomes, or missing clin- ical information were excluded. All the patients included met the following criteria: (1) aged 18-75 years; (2) hadn’t experienced local or systematic therapy before surgery; (3) with availability of the perioperative clinical data to calcu- late the S-GRAS score (including age, contrast CT image, TNM stage, clinical symptoms information Ki67 index); and (4) with availability of the follow-up radiological data to determine the disease status and survival. Finally, 39 patients were enrolled in this study.
Collection of clinical variables
We collected clinical characteristics of the patients, including PFS, OS, the radiologically confirmed relapse status, age, sex, ENSAT stage, resection status, preoperative symptoms, and Ki67%.
In this research, primary outcomes were: (1) PFS, defined as the interval between primary resection of ACC to the first radiologically identified progression, including disease relapse in patients after radical resection or pro- gressive and/or new lesions in patients with advanced
Pathologically proved ACC patients from Ruijin Hospital (n=187)
Four features were selected: original_shape_SurfaceArea logarithm_glszm_GrayLevelVariance wavelet.LHH_firstorder_Median wavelet.HLL_firstorder_Maximum
With available pre-operative raw contrast CT image. (117 excluded)
With available preoperative clinical data (including age, TNM stage, clinical symptoms information, Ki67 index). (1 excluded)
Suitable for surgery and hadn’t experienced local or systematic therapy before surgery. (21 excluded)
Randomly Separated Subsets
Subset 1
With available follow-up information. (9 excluded)
Subset 2
Subset 3
39 patients included
LASSO Regression
Subset 10
9 for Training
1 for Validation
a. ACC tumor identification
b. ACC tumor segmentation
c. Feature Extration
disease; (2) OS, defined as the time from primary resection of ACC to death. All patients were subject to regular follow-up assessments through radiological examinations every 3 months within the initial year post-surgery. Sub- sequently, these assessments were extended to every 6 months. The follow-up process continued until one of the following events occurred: death, loss to follow-up, or the most recent follow-up appointment. Follow-up was con- ducted through various means, such as outpatient visits, hospitalizations or telephone contact, to gather information about disease advancement, metastasis, and mortality. In the event of a patient’s demise, the date of death and the cause were recorded. The final follow-up was in April 2023.
Preoperative symptoms were defined either as hormone- related symptoms due to adrenal hormone excess such as in hypertension, Cushing syndrome or hirsutism, or as tumor- related symptoms such as abdominal pain. We sub- categorized patients by four different syndromes: Hyper- tension, Low Back Pain, Cushing’s and Aldosteronism.
The resection status was categorized into three types: R0, complete resection; Rx, cannot be assessed; R1, all mac- roscopic lesions are removed but microscopic margins are positive; R2, palliative surgery with gross lesions including the primary tumor, regional lymph nodes and involvement of macroscopic margins. The S-GRAS score was calculated according to following rules: age (<50 years = 0 point and ≥50 years = 1 point); pre-operation symptoms (No=0 point; Yes = 1 point); ENSAT stage (1 or 2 = 0 point; 3 = 1 point; 4 = 2 points); the resection status (R0 = 0 point; RX =1 point; R1 =2 points; R2 = 3 points); and Ki67% (0-9%=0 point; 10-19% = 1 point; ≥20% =2 points).
These generate 10 S-GRAS scores and four S-GRAS groups: 0-1, 2-3, 4-5, and 6-9 [7].
Radiomic feature extraction
The radiomic workflow is shown in Fig. 1. Patients who underwent contrast abdominal CT examination before initial ACC operation were included for analysis.
Image preprocessing and feature extraction were con- ducted using 3D Slicer (version 5.2.1). The segmentation process was carried out by a urologist with expertize in ACC and an experienced radiologist. To select robust fea- ture, several steps were performed. (1) A semi-automated approach was employed to draw the ROIs. The uppermost and lowermost slices of tumor were delineated manually. Subsequently, the “grow from seed” tool in 3D Slicer was employed to create the complete ROI. In instances where clear inconsistencies arose between the readers during this process, these discrepancies were resolved through agree- ment. Otherwise, the semi-automatically generated ROI mask was used for feature extraction. (2) As for pre- processing, the images were standardized by using “histo- gram matching” to eliminate the intensity variation and then resampled to a voxel size of 1 × 1 × 1 mm to standardize voxel spacing. Voxel intensity values were discretized using a fixed bin width of 25 HU. (3) 1413 features meet the Image Biomarker Standardization Initiative (IBSI) were extracted by “Pyradiomics” package. The image of 15 patients were re-evaluated by both two doctors twice. Both intra- and inter-doctor intraclass correlation coefficient (ICC) for each feature were calculated and the variables
with an ICC> 0.75 were included. In total, 1411 features were extracted from one CT image.
Construction of the radiomic index
Due to the small sample size, features were standardized, and the 10-fold cross validation least absolute shrinkage and selection operator (10-fold CV LASSO) cox regression model with “glmnet” R package was used to select sig- nificant features from high-dimensional data. The sig- nificant variables were then combined into a radiomic index (RDindex) by linearly adding up weighed by their respec- tive coefficients.
Index validation and comparison
The association between RDindex and PFS and OS was evaluated by Kaplan-Meier survival analysis. According to the optimal threshold of RDindex by using “survminer” package in R, patients were categorized into a high-risk group and a low-risk group, and difference between the two groups were analyzed by log-rank test.
Model construction and nomogram generation
Univariable cox regression was performed to select sig- nificant variables from age, sex, ENSAT stage, resection status, preoperative symptoms, Ki67% and RDindex. The significant variables identified in the univariable Cox regression analysis were subsequently subjected to the multivariable Cox regression model. Age and preoperative symptoms, which have been previously validated as valu- able predictors of prognosis, were also considered in the multivariable analysis [15].
The SGRAS score was not subjected to multivariable cox regression because it is an index constructed by part of these variables Given the small sample size, only significant and narrowly significant variables were integrated with RDin- dex to construct a nomogram. The SGRAS score or the ENSAT staging system alone was also used to generate nomograms.
Comparison between the models
The above-mentioned three nomograms (RDindex included, SGRAS score alone, and ENSAT stage alone) were com- pared. The incremental value of the RDindex was assessed with respect to calibration, discrimination, reclassification, and clinical usefulness. To compare the agreement between the predicted outcomes and the PFS, OS association, cali- bration curves were generated. To quantify the discrimina- tion performance, the Harrell concordance index (C-index) was measured. The Akaike information criterion (AIC) was
applied to avoid overfitting and anova analysis was applied to compare the goodness of fit. Finally, decision curve analysis was performed to determine the clinical usefulness of the nomograms by quantifying the net benefits at dif- ferent threshold probabilities. The statistical analysis was performed with R software, version 4.2.2, and RStudio, version 2023.3.1.446.
Results
Patients demographics
From September 2010 to August 2022, a total of 39 patients (19 male and 20 female) who met the inclusion criteria were included in this study (Table 1). Among them, 3 (7.69%) were classified as ENSAT stage 1, 14 (35.9%) as stage 2, 14 (35.9%) as stage 3, and 8 (20.5%) as stage 4. Most patients (82.5%) underwent RO resection. Among all patients, 28 (71.8%) experienced radiologically confirmed relapse. The median PFS was 20.4 months (interquartile range [IQR] 9.1-60.1 months), and the median follow-up period exten- ded for 52 months, IQR 16.8-49.2 months (Supplementary Table 1). As of the latest follow-up, disease progression was observed in 28 patients (71.8%), and 26 patients (66.7%) were confirmed to have deceased.
Radiomic features and clinical features
Radiomic features
Among the 1413 radiomic features extracted from contrast CT, 4 significant features (“original_shape_SurfaceArea”, “loga- rithm_glszm_GrayLevelVariance”, “wavelet.LHH_firstorder_ Median” and “wavelet.HLL_firstorder_Maximum”) were selected according to LASSO Cox regression after standardization.
RDindex was constructed by centralizing, standardizing and linearly combining them weighed by their respective coefficients.
RDindex= (“original_shape_SurfaceArea”-0.453)/ 0.0917*0.26+ (“logarithm_glszm_GrayLevelVariance”- 0.0384)/0.0078*0.011- (“wavelet.LHH_firstorder_Median” -0.011)/0.015*0.064+
(“wavelet.HLL_firstorder_Maximum”-31.02)/52.44*0.030
A Kaplan-Meier curve was constructed to assess the PFS based on the RDindex.
Under the optimal cutoff of -0.0937, patients were categorized into a high-risk group (RDindex ← 0.0937) and a low-risk group (RDindex ≥-0.0937). PFS, OS, recurrence rate and death rate in these two groups are shown in Table 2. Log-rank test revealed a statistically significant discrimination in PFS between the high- and low-risk
| Median (IQR)/n(%) | PFS | OS | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Univariate Cox Regression | Multivariate Cox Regression | Univariate Cox Regression | Multivariate Cox Regression | ||||||
| HR (95% CI) | P-Value | HR (95% CI) | P-Value | HR (95% CI) | P-Value | HR (95% CI) | P-Value | ||
| Age | 52.5 (38.0-60.0) | 1.00 (0.97-1.03) | 0.99 | 0.99 (0.95-1.04) | 0.77 | 1.01 (0.98-1.04) | 0.57 | 1.02 (0.97-1.06) | 0.484 |
| Sex | 0.94 (0.44-1.98) | 0.9 | / | / | 0.85 (0.39-1.87) | 0.68 | / | / | |
| Male | 19 (48.7) | / | / | / | / | / | / | / | / |
| Female | 20 (51.3) | / | / | / | / | / | / | / | / |
| ENSAT Stage | 2.29 (1.37-3.84) | 0.002 | 1.67 (0.97-2.88) | 0.06 | 2.40 (1.42-4.06) | 0.001 | 1.88 (1.03-3.40) | 0.04 | |
| Stage 1 | 3 (7.69%) | / | / | / | / | / | / | / | / |
| Stage 2 | 14 (35.9%) | / | / | / | / | / | / | / | / |
| Stage 3 | 14 (35.9%) | / | / | / | / | / | / | / | / |
| Stage 4 | 8 (20.5%) | / | / | / | / | / | / | / | / |
| Resection Status | 2.42 (1.34-4.40) | 0.004 | 3.60 (1.62-8.05) | 0.002 | 2.61 (1.45-4.73) | 0.002 | 3.00 (1.33-6.78) | 0.008 | |
| R0 | 32 (82.05%) | / | / | / | / | / | / | / | / |
| R1 | 4 (10.26%) | / | / | / | / | / | / | / | / |
| R2 | 3 (7.69%) | / | / | / | / | / | / | / | / |
| Preoperative Symptoms | / | / | / | / | / | / | / | / | / |
| No | 16 (41.03%) | 1.27 (0.59-2.76) | 0.54 | 0.57 (0.23-1.41) | 0.22 | 1.27 (0.56-2.87) | 0.57 | 0.73 (0.30-1.78) | 0.22 |
| Hypertension | 11 (28.21%) | 1.93 (0.80-4.65) | 0.15 | 1.39 (0.54-3.61) | 0.49 | 1.49 (0.57-3.89) | 0.41 | 1.16 (0.40-3.32) | 0.78 |
| Low Back Pain | 8 (20.51%) | 0.71 (0.24-2.07) | 0.53 | 0.08 (0.01-0.41) | 0.002 | 0.84 (0.26-2.76) | 0.77 | 0.37 (0.08-1.74) | 0.21 |
| Cushing's | 3 (7.69%) | 1.53 (0.33-7.17) | 0.59 | 0.38 (0.05-4.76) | 0.30 | 1.39 (0.36-5.36) | 0.64 | 0.68 (0.16-2.91) | 0.61 |
| Aldosteronism | 1 (2.56%) | 2.01 (0.25-16.2) | 0.51 | 0.48 (0.05-4.76) | 0.53 | 2.60 (0.32-21.2) | 0.37 | 0.45 (0.05-4.49) | 0.50 |
| Ki67% | 20.0 (10.0-32.5) | 1.02 (1.00-1.04) | 0.04 | 1.04 (1.01-1.08) | 0.005 | 1.02 (1.00-1.04) | 0.075 | 1.02 (1.00-1.05) | 0.099 |
| Radiation | 14 (35.90%) | 0.74 (0.34-1.65) | 0.47 | / | / | 0.86 (0.37-2.03) | 0.73 | / | / |
| Chemotherapy | 15 (38.46%) | 1.06 (0.50-2.26) | 0.88 | / | / | 0.98 (0.45-2.17) | 0.97 | / | / |
| Mitotane | 20 (51.28%) | 2.64 (1.14-6.09) | 0.023 | / | / | 2.10 (0.92-4.79) | 0.077 | / | / |
| SGRAS Score | / | 1.40 (1.15-1.70) | <0.001 | / | / | 1.39 (1.14-1.70) | 0.001 | / | / |
| 0-1 | 5 (12.82%) | / | / | / | / | / | / | / | / |
| 2-3 | 11 (28.21%) | / | / | / | / | / | / | / | / |
| 4-5 | 14 (35.90%) | / | / | / | / | / | / | / | / |
| 6-9 | 9 (23.08%) | / | / | / | / | / | / | / | / |
| RDIndex | -0.09 (-0.15-0.10) | 0.21 (0.04-0.98) | 0.05 | 0.05 (0.01-0.46) | 0.008 | 0.20 (0.04-1.07) | 0.06 | 0.03 (0.002-0.34) | 0.005 |
ENSAT Stage European network for the study of adrenal tumors, RDindex radiomics index, PFS progression free survival Median (IQR), Number (%)
| PFS | OS | |||||
|---|---|---|---|---|---|---|
| High Risk | Low Risk | Total | High Risk | Low Risk | Total | |
| No. of Patients | 19 | 20 | 39 | 19 | 20 | 39 |
| Survival (Month) | ||||||
| Median | 9.6 | 60.5 | 20.4 | 17 | 79 | 57.8 |
| Shortest | 1.2 | 1.4 | 1.2 | 1.5 | 20.4 | 1.5 |
| No. of Event | ||||||
| At 1 year | 9 (47.37%) | 5 (12.82%) | 14 (35.9%) | 3 (15.79%) | 0 (0.00%) | 3 (7.69%) |
| At 2 years | 14 (73.68%) | 7 (17.95%) | 21 (53.85%) | 10 (52.63%) | 1 (5.00%) | 11 (28.21%) |
| At 3 years | 14 (73.68%) | 7 (17.95%) | 21 (53.85%) | 12 (63.16%) | 1 (5.00%) | 13 (33.33%) |
Number (%)
Strata + Rdindex=HighRisk + Rdindex=LowRisk
Strata + Rdindex=HighRisk + Rdindex=LowRisk
1.00
1.00
Progression Free Suvival Probability
0.75
0.75
Overall Suvival Probability
0.50
0.50
0.25
0.25
p = 0.0025
p = 0.0023
0.00
0.00
0
25
50
75
100
0
50
100
150
Time
Time
Number at risk
Number at risk
Rdindex=HighRisk
19
4
Strata
2
1
0
Rdindex=HighRisk
19
5
1
0
Strata
Rdindex=LowRisk
20
14
11
3
0
Rdindex=LowRisk
20
17
1
0
0
25
50
75
100
0
50
100
150
Time
Time
groups (p=0.0025), and OS between these two groups (p=0.0023) (Fig. 2).
Clinical features
Progression free survival
Univariate Cox regression analysis identified the ENSAT stage, resection status and Ki67% as significant hazard fea- tures, with a hazard ratio (HR) of 2.29 (95% CI 1.34-3.84, p =0.002), 2.42 (95% CI 1.34-4.40, p = 0.004), and 1.02 (95% CI 1.00-1.04, p = 0.04) respectively (Table 1). Whe- ther patients exhibited preoperative symptoms or not, and the different subtypes of preoperative symptoms, are not con- sidered as significant variables affecting the prognosis (No symptoms HR 1.27, 95%CI 0.59-2.76, p =0.54,
Hypertension HR 1.93, 95%CI 0.80-4.65, p = 0.54, Low Back Pain HR 0.71, 95%CI 0.24-2.07, p=0.53, Cushing’s HR 1.53, 95%CI 0.33-7.17, p=0.59, Aldosteronism HR 2.01, 95%CI 0.25-16.2, p=0.51). The postoperative adjunctive treatments, including radiation and chemotherapy, were not considered significant variables, with HR values of 0.74 (95% CI 0.34-1.65, p=0.47) and 1.06 (95% CI 0.50-2.26, p =0.88), respectively. However, Mitotane was identified as a significant hazard variable, with an HR of 2.64 (95% CI 1.14-6.09, p = 0.023). Preoperative symptoms were not regarded as a significant variable (HR 1.27, 95% CI 0.59-2.76, p = 0.54). The SGRAS model was demonstrated as having a reliable prediction ability, with an HR of 1.40 (95% CI 1.15-1.70, p<0.001). Additionally, the combined RDindex was found to be a significant protective factor, with an HR of 0.21 (95% CI 0.04-0.98, p = 0.05).
a
RDindex
S-GRAS
ENSAT
Points
0
10
20
30
40
50
60
70
80
90
100
Points
0
10
20
30
40
50
60
70
80
100
Points
0
10
20
30
40
50
60
70
80
90
100
ENSAT
2
4
1
3
Rdindex
SGRAS
1
3
5
7
2
4
0
2
4
6
B
ENSAT
1
3
1.2
1
0.8
0.6
0.4
02
·
02
-0.4
0.6
OperationType
N
·
2
Total Points
0
Total Points
10
20
30
40
50
60
TO
80
90
100
0
10
20
30
40
50
60
70
80
90
100
KJ67%
0
10
20
30
40
50
60
70
80
Total Points
1 year survival
1 year survival
@
20
40
00
100
120
140
160
100
200
220
240
260
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.9
0.8
0.7
0.6
0.5
0.4
1 year survival
0.9
0.8
0.7 0.6 0.5 0.4 0.3 0.2 0.1
2 year survival
2 year survival
0.8
2 year survival
0.7
0.6
0.5
0.4
0.3
02
0.1
D.B
0.7
0.6
0.5
0.4
0.3
02
0.1
09
0.8
0.7
0.6 05 04 03 02
0.1
5 year survival
0.7 06 05 04 03 02 01
5 year survival
D.7
0.1
5 year survival
06
0.5
0.4
0.3
02
07
0.6
9.5
0.4
03
02
0.1
b
One year PFS Calibration Curve
Two years PFS Calibration Curve
Five years PFS Calibration Curve
9
9
9
T
0
8
0
Actual DFS (Proportion)
Actual DFS (Proportion)
Actual DFS (Proportion)
8
8
8
8
8
8
a
영
0
RDindex
RDindex
RDindex
SGRAS
SORAS
SGRAS
0
ENSAT
8
ENSAT
0
ENSAT
0.0
02
0.4
0.6
0.8
1.0
0.0
02
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
Predicted Probability
Predicted Probability
Predicted Probability
C
RDindex
S-GRAS
ENSAT
Points
0
10
20
30
40
50
70
90
100
Points
D
10
20
30
40
50
TO
50
90
100
Points
D
10
20
30
40
50
70
80
BO
100
ENSAT
2
4
1
3
SORAS
1
3
5
7
ENSAT
2
4
Rdindex
Di
4
&
1
12
3
1
0.8
0.6
0.4
2
8
02
0
02
-0.4
-0.6
OperationType
9
2
Total Points
0
Total Points
10
20
30
40
50
60
00
90
100
D
10
20
30
40
50
60
70
90
100
P067%
D
10
N
0
4
0
0 70 0
Total Points
2 year survival
2 year survival
a
20
60
80
100
120
140
160
180
200
0.9
08
0.7
0.6
0.5
0.4
0.7
0.5
2 year survival
0.9
07 0.60.50.40.302 0.1
3 year survival
3 year survival
0.9
0.8
0.7
0.6
0.5
0.4
03
3 year survival
09
0.7
0.6
0.5
9.4
0.9
0.8 0.7 0.60 50.40 30.2 0.1
5 year survival
0.9
0.8 0.7 0.60.50.40.30.2 0.1
5 year survival
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
5 year survival
o.a
0.7
0.6
0.5
0.4
0.3
0.2
0.1
d
Two year OS Calibration Curve
Three years OS Calibration Curve
Five years OS Calibration Curve
8
-
m
TT
T
2
II
TT
I
IT
L
m
2
I
r
TT
m
T
T
×
2
8
&
Actual DFS (Proportion)
Actual DFS (Proportion)
Actual DFS (Proportion)
0
S
0.6
3
8
0.4
3
3
3
RDindex
RDindex
RDindex
SGRAS
SGRAS
O
8
I
SGRAS
ENSAT
ENSAT
ENSAT
0.0
02
0.4
0.6
0.8
1.0
0.0
02
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
Predicted Probability
Predicted Probability
Predicted Probability
Overall survival
Similarly, the ENSAT stage and resection status were identified as significant hazard features, with HR of 2.40 (95% CI 1.42-4.06, p =0.001) and 2.61 (95% CI 1.33-6.78, p =0.008), respectively, and the syndrome was regarded as not significant (Table 1). Ki67% was a marginally significant variable (HR 1.02, 95% CI 1.00-1.04, p = 0.075). Additionally, mitotane was con- sidered a marginally significant hazard variable for OS, with an HR of 2.10 (95% CI 0.92-4.79, p = 0.077). Notably, the SGRAS model continued to demonstrate predictive performance, with an HR of 1.39 (95% CI 1.14-1.70, p = 0.001). Although the p-value for RDin- dex was 0.06 (HR 0.20, 95% CI 0.04-1.07), it is con- sidered a marginally significant hazard variable and was
subsequently included in the multivariable Cox regres- sion model.
Multivariate Cox regression
To further evaluate these features, a multivariate cox regression analysis was applied. As the SGRAS score was generated based on these variables, it was not subjected to multivariable Cox regression to avoid collinearity.
In Fig. 3a From left to right exhibits the RDindex nomogram, S-GRAS and ENSAT nomogram. Sum the points achieved for each of the risk factors. Locate the final sum on the Total Point axis. Draw a line straight down to find the patient’s probability of PFS. (b) From left to right exhibits the calibration curve of three nomogram to predict 1-year, 2-year and 5-year PFS respectively. (c) From left to
right exhibits the RDindex nomogram, S-GRAS and ENSAT nomogram. Sum the points achieved for each of the risk factors. Locate the final sum on the Total Point axis. Draw a line straight down to find the patient’s probability of OS. (d) From left to right exhibits the calibration curve of three nomogram to predict 2-year, 3-year and 5-year OS respectively.
Progression free survival
The resection status (HR 3.60, 95%CI 1.62-8.05, p = 0.002), Ki67% (HR 1.04, 95%CI 1.01-1.08, p = 0.005) and RDindex (HR 0.05, 95%CI 0.01-0.46, p = 0.008) were regarded as significant variables. — Given the limited sample size, ENSAT stage (HR 1.67, 95%CI 0.97-2.88, p= 0.06) was narrowly significant.
Overall survival
Similarly, the resection status (HR 3.00, 95%CI 1.33-6.78, p=0.008) and RDindex (HR 0.03, 95%CI 0.002-0.34, p = 0.005) were identified as significant hazard features. ENSAT stage showed a significantly predictive perfor- mance with the HR of 1.88 (95%CI 1.03-3.40, p = 0.04). The Ki67% (HR 1.02, 95%CI 1.00-1.05, p = 0.099) was marginally significant.
Nomograms and model performance
All these four significant or narrowly significant variables (ENSAT stage, resection status, Ki67% index and RDin- dex) were merged to construct a nomogram (Fig. 3).
Progression free survival
Figure 3a shows the PFS nomograms of the RDindex model and SGRAS model. According to the calibration curve (Fig. 3b), the value of the three models in predicting the 1-, 2- and 5-year tumor recurrence rates was compared. Their calibration performance was similar and showed a good predictive calibration, especially in predicting 2-year PFS. C-index, AIC and the p-value (compared with RDin- dex nomogram) calculated by Anova test of these three models (RDindex nomogram, SGRAS score and ENSAT stage) are listed in Table 3. Anova test showed that RDin- dex had the highest C-index (0.75) and the lowest AIC (161) as compared either with the SGRAS Score (C-index 0.68 and AIC 164, P=0.030) or the ENSAT Stage (C- index 0.67 and AIC 165, p = 0.025). Finally, a decision curve analysis was performed, and the result showed that the RDindex model performed better net benefit than either the SGRAS model or the ENSAT model across the most range of threshold probabilities. (Fig. 4)
| PFS Model | OS Model | |||||
|---|---|---|---|---|---|---|
| C-Index | AIC | P Value* | C-Index | AIC | P Value* | |
| Radiomics | 0.75 | 161 | / | 0.78 | 133 | / |
| nomogram | ||||||
| S-GRAS | 0.68 | 164 | 0.039 | 0.72 | 141 | 0.003 |
| Scoreª | ||||||
| ENSATª | 0.67 | 165 | 0.025 | 0.73 | 140 | 0.006 |
AIC akaike information criteria
ªCompared to Radiomics nomogram
Figure 4 is the decision curve analysis for each model (4a is for PFS, 4b is for OS). The y-axis measures the net benefit. The net benefit was calculated by summing the benefits (true positive results) and subtracting the harms (false positive results), weighting the latter by a factor related to the relative harm of an undetected cancer com- pared with the harm of unnecessary treatment. The RDindex model had the highest net benefit.
Overall survival
Figure 3c displays the nomogram for predicting OS. Unlike the PFS model, this nomogram compares the predictive abilities of these three models for 2-, 3-, and 5-year OS. In Fig. 3d, their calibration curves are compared. These three models have similar calibration curves and exhibit good predictive calibration, especially when predicting 3-year OS. Table 3 also presents the C-index, AIC, and p-value for these three models in predicting OS. The results of the ANOVA test indicated that the RDindex model exhibited the highest C-index at 0.78 and the lowest AIC score of 133. In comparison, the SGRAS model had a C-index of 0.72 and an AIC score of 141 (p=0.003), while the ENSAT model had a C-index of 0.73 and an AIC score of 140 (p=0.006). Similarly, the DCA revealed that when predicting OS, the RDindex model consistently demon- strated a superior net benefit compared to both the SGRAS model and the ENSAT model across a wide range of threshold probabilities (Fig. 4).
Discussion
ACC is a rare neoplasm characterized by a high risk of recurrence after radical resection. Our research findings align with previous studies, revealing that within a 2-year period, 53.85% ACC patients experienced disease recur- rence or progression. Furthermore, over a 5-year timeframe, a substantial proportion (74.36%) of ACC patients encountered recurrence or progression. These outcomes
a
0.4
Net Benefit
0.2
- RDindex Model
··· ENSAT Model
— SGRAS Model
- · All
None
0.0-
-0.2
0.25
0.50
0.75
1.00
b
Risk Threshold
0.3
0.2
Net Benefit
RDindex Model
0.1
··· ENSAT Model
— SGRAS Model
- . All
… None
0.0
-0.1-
0.00
0.25
0.50
0.75
1.00
Risk Threshold
underscore the aggressive behavior of ACC and emphasize the importance of developing effective prognostic strategies and therapeutic interventions to improve patient outcomes [16]. The ACC prognosis prediction model remains further explore as there was no previous study incorporate radiomic features into their predictive model in terms of ACC, although some have reported the construction of models based on clinicopathological features to predict prognosis. Kim et al. [17] constructed a nomogram to predict recurrence-free survival by selecting five clinical factors
including the tumor size, nodal status, T stage, capsular infiltration and hormone syndrome. A retrospective study including 319 patients [8] reported Ki67 index as the single most important variable to predict recurrence in ACC patients with R0 resection. A recent study by Rossella et al. [18] demonstrated a strong correlation between 18F-FDG uptake and Ki67 index. Several studies [2-4] have validated the combination of S-GRAS score with these factors as a promising model to provide urologists with promising instructions. However, none of these studies reported the
use of radiomics as a replacement for the preoperative TNM stage in predicting the prognosis of ACC patients. To the best of our knowledge, this study represents the first attempt to utilize radiomic analysis for predicting prognosis in ACC.
Radiomics has emerged as a highly promising and dependable field in the realm of cancer diagnosis, show- casing its versatility and effectiveness in a multitude of domains within cancer prognosis. By harnessing sophisti- cated image processing methodologies and machine learn- ing algorithms, radiomics empowers the extraction of a vast array of concealed insights from medical images, thereby augmenting the precision of diagnostic procedures, prog- nostic assessments, and evaluations of treatment response. Its widespread implementation spans diverse applications in such as lung cancer diagnosis, prognostic prediction of PFS in early-stage lung cancer patients, classification of distinct subtypes of glioblastoma, and prognostication of distant metastasis in hepatocellular carcinoma [12, 14, 19-21].It is also widely used in the field of kidney cancer [22]. But the application of radiomics in ACC related diagnosis or prognosis prediction hasn’t been published yet.
In this high-dimensional limited-sample sized study, we had performed 4 steps to ensure the stability and robustness of our research. (1) A semi-automated approach was employed to draw the ROIs, both the urologist and the radiologist only delineate the several uppermost and low- ermost slices of tumor, then the “grow from seed” tool in 3D Slicer was employed to create the complete ROI. The inconsistencies were solved by consensus. (2) The “histo- gram matching” were applied during the image preproces- sing to eliminate the intensity variation and then resampled to a voxel size of 1 × 1 × 1 mm to standardize voxel spacing. Voxel intensity values were discretized using a fixed bin width of 25 HU to mitigate image noise and standardize intensities, thereby ensuring consistent intensity resolution across all tumor images. (3) 1413 features meet the IBSI were extracted by “Pyradiomics” package. To further enhance the robustness of this research, the urologist and radiologist in charge of image reading reevaluated the CT images of 15 patients originally assessed by the urologist. We calculated both the intra- and inter-doctor ICC for each feature. Variables with an ICC>0.75 were included. (4) 1411 features were included into the following steps. These features were screened by a 10-fold cross validation LASSO regression. This involved splitting the dataset into 10 parts, using nine parts for model training and one part for vali- dation, and repeating this process 10 times to minimize bias to the best of our ability. The cross-validation LASSO regression is regarded as a reliable algorithm to select sig- nificant feature and to construct model in a high- dimensional low-sample-sized study and has also been applied in previous study to enhance model’s robustness
and reliability [23]. Those features exhibited inconsistency among patients and lacked significance were subsequently filtered out, retaining only the most robust and statistically significant four features for subsequent modeling.
Finally, a total of 1411 features from the contrast CT images were retained for further model construction. Using LASSO regression, we identified four significant features, including two derived from the original imaging data. The feature “original_shape_SurfaceArea” was found to be closely associated with tumor size, providing valuable information about the tumor’s physical dimensions. The feature “logarithm_glszm_GrayLevelVariance” was found to reflect the tumor texture characteristics. Notably, tumors exhibiting calcification and necrosis tended to exhibit higher values of “logarithm_glszm_GrayLevelVariance.” These two features provide additional information regarding tumor behavior, surpassing the limited scope of the conventional T stage, which primarily considers tumor size and neighbor- ing structures. The remaining two features were computed using Fourier transform and Wavelet transform techniques, rendering them more abstract in terms of their specific clinical interpretations. However, as mathematic features calculated by decomposing images into different levels or layers, they may not have direct clinical meanings, but still reflect important image characteristics.
Given that radiomic features primarily reflect the loca- lized characteristics of the tumor, their predictive power can be enhanced by integrating them with other relevant clinical variables. Thus, we have generated a prognosis prediction model by incorporating RDindex, ENSAT stage, resection status and Ki67% index. To be mentioned that, although the mitotane was found to be a significant factor in the uni- variate regression, we didn’t incorporate it into the model construction. That was because Mitotane was found to be a risk factor, rather than a protective factor, in both univariate regression analyses for PFS and OS (HR 2.64, HR 2.10, respectively), which is contrary to what was expected. One possible reason for this unexpected result could be that Mitotane was not available on the mainland of China during the postoperative and follow-up periods. Therefore, only patients with more severe conditions would have obtained it through special channels, introducing bias into the analysis. In addition, it’s important to note that this study does not primarily focus on treatment regimens. Currently, there is a lack of full standardization and consensus on drug therapies for ACC, such as the timing and duration of Mitotane use. Therefore, even if these factors were considered, they might introduce bias into the analysis due to their non- standardized nature.
The new model showed better performance than the S-GRAS score or ENSAT stage system alone in predicting the prognosis of ACC patients according to C-Index, AIC and Anova analysis. C-index of the ENSAT model was
0.67, which is consistent with 0.67 and 0.665 reported in two previous studies [7, 24], confirming the credibility and validity of our research endeavors.
The new model described in this study provides a novel approach for prognosis prediction of ACC and paves the way for future exploring effective therapeutic approaches for ACC. In current clinical practice, TNM staging relies primarily on tumor size, adjacent organ involvement, the presence of positive lymph nodes and distant metastasis, while radiomics offers the potential to provide not only these conventional factors but also additional hidden information that is imperceptible to the naked eye. Fur- thermore, the extraction of features using computer algo- rithms is less subjective compared with conventional methods. Looking ahead, there is a prospective future where radiomics could potentially supplant the simplistic TNM staging system, thus offering a more comprehensive and accurate assessment of tumor characteristics and prognosis.
This study has several limitations that warrant con- sideration. Firstly, the sample size was relatively small, consisting of only 39 patients, which can be attributed to the rarity of the disease. Additionally, the prediction model developed in this study needs external validation to ensure its generalizability and reliability. Although internal cross- validation and AIC analysis were conducted to address multicollinearity, an additional validation using an inde- pendent dataset is necessary. Secondly, only the cox regression model was constructed in this research. A further machine learning or deep learning-based model remain to be constructed. However, there are some research reported that the regression is sufficient and performs better than some kinds of machine learning algorithms [23]. A further comparison remains to be made. Lastly, all the images used in this study were reconstructed using the IR algorithm, ensuring that the results are not influenced by the recon- struction method. Nevertheless, future studies should vali- date how this model performs on images reconstructed using the FBP algorithm. Additionally, considering the variation in radiation doses among patients, with a median CTDI volume value of 10.2 (ranging from 7.6 to 12.4, Supplementary Table 1), image preprocessing methods were employed during feature extraction to minimize potential differences arising from varying radiation doses. However, inherent biases may still be present.
Conclusions
The integration of radiomic features extracted from ACC contrast CT images with the S-GRAS model has demon- strated the potential to enhance the prognosis prediction in terms of PFS.
Data availability
The datasets generated and analyzed during the current study are available from the corresponding authors on rea- sonable request.
Supplementary information The online version contains supplemen- tary material available at https://doi.org/10.1007/s12020-023-03568-4.
Acknowledgements Guarantor: The scientific guarantor of this pub- lication is J.Z. Methodology: Retrospective, prognostic study, per- formed at one institution.
Author contributions All authors contributed to the study conception. Methodology was designed by J.L., W.L. and J.Z. Software and Formal analysis were performed by J.L. Resources was prepared by J.D., D.X. and J.Z. Investigation was conducted by J.L. and J.Z. Data Curation was performed by J.L., W.L., L.Y., J.X. and J.Z. The first draft of the manuscript was written by JL. and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.
Compliance with ethical standards
Conflict of interest The authors declare no competing interest.
Ethical approval Institutional Review Board approval was obtained.
Informed consent Informed consent was obtained from all patients and/or legal guardians in this study.
References
1. T. Else, A.C. Kim, A. Sabolch et al. Adrenocortical carcinoma. Endocr. Rev. 35, 282-326 (2014). https://doi.org/10.1210/er. 2013-1029
2. J. Pura, M. Dinan, S. Reed et al. Treatment patterns and outcomes for patients with adrenocortical carcinoma associated with hospital case volume in the United States. Ann. Surg. Oncol. 21, 3509-3514 (2014). https://doi.org/10.1245/s10434-014-3931-z
3. T.M. Kerkhofs, R.H. Verhoeven, H.J. Bonjer et al. Surgery for adrenocortical carcinoma in The Netherlands: analysis of the national cancer registry data. Eur. J. Endocrinol. 169, 83-89 (2013). https://doi.org/10.1530/EJE-13-0142
4. M. Terzolo, A.E. Baudin, A. Ardito et al. Mitotane levels predict the outcome of patients with adrenocortical carcinoma treated adjuvantly following radical resection. Eur. J. Endocrinol. 169, 263-270 (2013). https://doi.org/10.1530/EJE-13-0242
5. M. Fassnacht, S. Johanssen, M. Quinkler et al. Limited prognostic value of the 2004 international union against cancer staging classification for adrenocortical carcinoma: proposal for a revised TNM classification. Cancer 115, 243-250 (2009). https://doi.org/ 10.1002/cncr.24030
6. T. Else, A.R. Williams, A. Sabolch et al. Adjuvant therapies and patient and tumor characteristics associated with survival of adult patients with adrenocortical carcinoma. J. Clin. Endocrinol. Metab. 99, 455-461 (2014). https://doi.org/10.1210/jc.2013-2856
7. Y.S. Elhassan, B. Altieri, S. Berhane et al. S-GRAS score for prognostic classification of adrenocortical carcinoma: an interna- tional, multicenter ENSAT study. Eur. J. Endocrinol. 186, 25-36 (2022). https://doi.org/10.1530/EJE-21-0510
8. F. Beuschlein, J. Weigel, W. Saeger et al. Major prognostic role of Ki67 in localized adrenocortical carcinoma after complete resec- tion. J. Clin. Endocrinol. Metab. 100, 841-849 (2015). https://doi. org/10.1210/jc.2014-3182
9. A. Berruti, M. Fassnacht, H. Haak et al. Prognostic role of overt hypercortisolism in completely operated patients with adrenocor- tical cancer. Eur. Urol. 65, 832-838 (2014). https://doi.org/10. 1016/j.eururo.2013.11.006
10. M. Fassnacht, O.M. Dekkers, T. Else 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. 179, G1-G46 (2018). https://doi.org/10.1530/EJE-18- 0608
11. W. Lin, J. Dai, J. Xie et al. S-GRAS score performs better than a model from SEER for patients with adrenocortical carcinoma. Endocr. Connect 11, e220114 (2022). https://doi.org/10.1530/EC- 22-0114
12. Y. Huang, Z. Liu, L. He et al. Radiomics signature: a potential biomarker for the prediction of disease-free survival in early-stage (I or II) Non-Small Cell Lung Cancer. Radiology 281, 947-957 (2016). https://doi.org/10.1148/radiol.2016152234
13. Y. Yu, Y. Tan, C. Xie et al. Development and validation of a preoperative magnetic resonance imaging radiomics-based sig- nature to predict axillary lymph node metastasis and disease-free survival in patients with early-stage breast cancer. JAMA Netw. Open 3, e2028608 (2020). https://doi.org/10.1001/jamanetw orkopen.2020.28086
14. L. Jiang, C. You, Y. Xiao et al. Radiogenomic analysis reveals tumor heterogeneity of triple-negative breast cancer. Cell Rep. Med. 3, 100694 (2022). https://doi.org/10.1016/j.xcrm.2022. 100694
15. R. Libé, I. Borget, C.L. Ronchi et al. Prognostic factors in stage III-IV adrenocortical carcinomas (ACC): an European Network for the Study of Adrenal Tumor (ENSAT) study. Ann. Oncol. 26, 2119-2125 (2015). https://doi.org/10.1093/annonc/mdv329
16. K.Y. ilimoria, W.T. Shen, D. Elaraj et al. Adrenocortical carci- noma in the United States. Cancer 113, 3130-3136 (2008). https:// doi.org/10.1002/cncr.23886
17. Y. Kim, G.A. Margonis, J.D. Prescott et al. Nomograms to predict recurrence-free and overall survival after curative resection of
adrenocortical carcinoma. JAMA Surg. 151, 365-373 (2016). https://doi.org/10.1001/jamasurg.2015.4516
18. R. Libé, A. Pais, F. Violon et al. Positive correlation between 18F- FDG uptake and tumor-proliferating antigen Ki-67 expression in adrenocortical carcinomas. Clin. Nucl. Med 48, 381-386 (2023). https://doi.org/10.1097/RLU.0000000000004593
19. J.E. Park, H.S. Kim, S.Y. Park et al. Prediction of core signaling pathway by using diffusion- and perfusion-based MRI radiomics and next-generation sequencing in isocitrate dehydrogenase wild- type glioblastoma. Radiology 294, 388-397 (2020). https://doi. org/10.1148/radiol.2019190913
20. G.W. Ji, F.P. Zhu, Q. Xu et al. Radiomic features at contrast- enhanced CT predict recurrence in early-stage hepatocellular carcinoma: a multi-institutional study. Radiology 294, 568-579 (2020). https://doi.org/10.1148/radiol.2020191470
21. H. Tibermacine, P. Rouanet, M. Sbarra et al. Radiomics modelling in rectal cancer to predict disease-free survival: evaluation of different approaches. Br. J. Surg. 108, 1243-1250 (2021). https:// doi.org/10.1093/bjs/znab191
22. G. Yang, P. Nie, L. Yan et al. The radiomics-based tumor het- erogeneity adds incremental value to the existing prognostic models for predicting outcome in localized clear cell renal cell carcinoma: a multicenter study. EJNMMI 49, 2949-2959 (2022). https://doi.org/10.1007/s00259-022-05773-1
23. S. Ghezzo, P. Mapelli, C. Bezzi et al. Role of [68Ga] Ga-PSMA- 11 PET radiomics to predict post-surgical ISUP grade in primary prostate cancer. EJNMMI 50, 2548-2560 (2023). https://doi.org/ 10.1007/s00259-023-06187-3
24. J. Lippert, B. Altieri, B. Morrison et al. Prognostic role of targeted methylation analysis in paraffin-embedded samples of adreno- cortical carcinoma. J. Clin. Endocrinol. Metab. 107, 2892-2899 (2022). https://doi.org/10.1210/clinem/dgac470
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.