Prognostic Value of the Radiomics-Based Model in the Disease-Free Survival of Pretreatment Uveal Melanoma: An Initial Result : Journal of Computer Assisted Tomography

Secondary Logo

Journal Logo

Neuroimaging: Head and Neck

Prognostic Value of the Radiomics-Based Model in the Disease-Free Survival of Pretreatment Uveal Melanoma: An Initial Result

Su, Yaping MD∗,†; Xu, Xiaolin MD†,‡,§; Wang, Fang MD; Zuo, Panli MD; Chen, Qinghua MD∗,†; Wei, Wenbin MD, PhD†,‡,§; Xian, Junfang MD, PhD∗,†

Author Information
Journal of Computer Assisted Tomography 47(1):p 151-159, 1/2 2023. | DOI: 10.1097/RCT.0000000000001384


Uveal melanoma (UM), the most common primary intraocular malignant tumor, mainly has a greater than 50% propensity for metastases and most commonly metastasizes to the liver.1,2 Surgical enucleation and definitive radiotherapy have improved local control. Despite these therapeutic efforts, up to 50% of patients eventually die of this disease.3–5 Therefore, an accurate pretreatment risk assessment for patients with UM is urgently needed so that aggressive therapeutic strategies can be implemented for the treatment of high-risk individuals.

At present, the techniques for assessing the prognosis of UM have evolved from assessing simple physical features to slightly more sophisticated ones, such as chromosomal mutations and genetic mutations.6,7 Several clinical and histological features, such as larger tumor size, extraocular extension, and epithelioid cell type, can predict poor prognosis; however, the accuracy of the assessment needs to be further improved.6 Previous studies have identified that patients with loss of chromosome 3 or disomy 3 with amplification of chromosome 8q have poor prognosis.1 More recently, studies report that gene mutations in GNAQ, GNA11, BAP1, and SF3B1 are associated with UM prognosis.2,8,9 Reportedly, these methods have enhanced the accuracy of prognostic testing and precision medicine in UM. Owing to the high cost and complex processes of molecular biological techniques, these techniques are not currently used in a clinical setting. Besides, physical features rely on postoperative pathology. Accordingly, a new and cost-effective method is warranted for prognostic evaluation.

To date, multiparametric magnetic resonance imaging (MRI) protocols, including dynamic contrast-enhanced imaging and diffusion-weighted imaging, have been generally used for the predication of tumor progression. Kamrava et al10 have reported that higher Ktrans (transfer constant from the blood plasma into the extracellular extravascular space) and higher Ve (extracellular extravascular volume fraction) values correlate with higher monosomy 3. Foti et al11 identified that both pretreatment and an early change in apparent diffusion coefficient value may predict the treatment response of UM. The MRI scan times for dynamic contrast-enhanced imaging and diffusion-weighted imaging are sufficiently long such that eye movements during the scan result in image artifacts and air-bone interface result in image distortion, which lead to inaccurate results. Therefore, a noninvasive and highly predictive alternative imaging method is warranted.

Radiomics is the high-throughput extraction of a quantity of imaging features from clinically acquired images and the conversion of these medical images into mineable high-dimensional data. The potential benefit of radiomics has been highlighted in different clinical applications, such as cancer detection and classification, phenotypic subtype, treatment response monitoring, and prediction of clinical outcomes.12–17 Previous study has shown that radiomics analysis represents a promising tool in separating UM from other intraocular masses with the area under curve range from 0.775 to 0.877 in the different models.18 However, a model for assessing the risk of disease-free survival (DFS) in UM has yet to be reported.

Thus, the aim of this study was to develop an MRI-based radiomics model for the prediction of the DFS of patients with UM before pretreatment. We also incorporated a radiomics score (rad-score) with visual MRI findings to build a visual nomogram with greater accuracy for individualized patient stratification.



This retrospective study was approved by our institutional review board. Informed consent was waived for all patients. All patient data and personal information were anonymized before analysis. The study's inclusion criteria were as follows: (a) patients who underwent enucleation or resection for UM verified via histopathology; (b) those who underwent MRI before treatment (including T1-weighted images [T1WI], T2-weighted images, [T2WI], and contrast-enhanced T1-weighted images [CET1WI]); (c) those with no prior history of cancer treatment; (d) those without distant metastases at the time of diagnosis; or (e) those with complete follow-up history. The exclusion criteria were as follows: (a) patients without contrast-enhanced and noncontrast MRI before treatment; (b) those with poor-quality images (eg, significant motion and susceptibility artifacts) or with corrupted digital imaging and communications in medicine (DICOM) image files; (c) those with a partial or no follow-up history; or (d) those with a history of previous or synchronous malignant tumors. Of the 119 patients with UM confirmed via histopathology who were treated from March 2007 to September 2017 at our hospital, 5 were excluded owing to missing MRI data, 6 were excluded owing to corrupted DICOM image files, and 23 were excluded owing to insufficient follow-up data. Therefore, the total study sample was 85 patients. These patients were randomly assigned to the training cohort (n = 60; median age, 44 years) and validation cohort (n = 25; median age, 43 years).

Follow-up and Survival Endpoint

Patients were followed up every 1 to 3 months within a year of treatment, every 3 to 6 months in year 2, 4 to 8 months in years 3 to 5, and annually thereafter. After enucleation, patients received follow-up care at their local hospital unless they lived close to our center. Other than the first outpatient visit, patients received follow-up care in an outpatient clinic or via a telephonic interview. The follow-up appointment included liver function tests, clinical abdomen computed tomography or MRI, chest radiography or computed tomography, spinal MRI, and a review of their ocular condition. The endpoint of this study was DFS. The study's endpoint, DFS, was defined as the time from the time of surgery until the date of the first locoregional recurrence, distant metastasis, death, or the last follow-up visit.19 The follow-up period range was 6 to 125 months after treatment. Demographic data were obtained at baseline for each patient with UM, including age, sex, and follow-up data (follow-up duration and survival), and were obtained from their medical records. The tumor was staged according to the eighth edition of the American Joint Committee on Cancer Staging Manual for UM.20 Patients were divided into 2 groups: the early-stage group consisting of patients with T1 and T2 stage tumors and the advanced stage group including patients with T3 and T4 stage tumors (Table 1).

TABLE 1 - Clinical Characteristics and MRI Findings of Patients With UM in the Training and Validation Cohorts
Characteristic Training Cohort (n = 60) Validation Cohort (n = 25) P
Age 0.552
 Median (range) 44 (19–61) 43 (18–65)
Gender, n (%) 0.812
 Male 34 (57) 13 (52)
 Female 26 (43) 7 (48)
Location, n (%) 0.226
 Ciliary body 0 0
 Anterior without ciliary infiltration 8 (13) 0
 Anterior with ciliary infiltration 6 (10) 2 (8)
 Posterior 33 (55) 16 (64)
 Crossing the equator 13 (22) 7 (28)
Shape, n (%) 0.165
 Placoid shaped 6 (10) 1 (4)
 Lentiform shaped 11 (18) 7 (28)
 Mound shaped 25 (42) 13 (52)
 Mushroom shaped 18 (30) 4 (16)
Margin, n (%) 0.165
 Well defined 49 (82) 24 (96)
 Blurred margin 11 (18) 1 (4%)
Signal intensity, n (%)
 T1WI 0.805
 Hypointensity 0 0
 Hyperintensity 38 (63) 17 (68)
 Isointensity 22 (37) 8 (32)
 T2WI 0.811
 Hypointensity 36 (60) 14 (56)
 Hyperintensity 0 0
 Isointensity 24 (40) 11 (44)
Height, mm 0.492
 Median (range) 8 (2–12) 7 (3–10)
Basal diameter, mm 0.437
 Median (range) 11 (5–18) 11 (5–15)
T stage, n (%) 0.531
T1, T2 34 (57) 16 (64)
T3, T4 26 (43) 9 (36)
Degree of enhancement, n (%) 0.864
 Mild 47 (78) 20 (80)
 Non-mild 14 (22) 5 (20)
Homogeneity of enhancement, n (%) 0.204
 Homogenous 37 (62) 19 (76)
 Heterogeneous 23 (38) 6 (24)
Retinal detachment, n (%) 0.154
 Absence 19 (32) 13 (52)
 Presence 41 (68) 12 (48)
Mean DFS time, mo 0.169
 Median (range) 56 (6–122) 68 (13–125)

Image Acquisition and Segmentation

Figure 1 shows the radiomics workflow. All MRI scans were obtained using a 1.5-T MRI scanner (n = 46, Signa Highspeed; GE Healthcare, Milwaukee, Wis) or a 3.0-T MRI scanner (n = 23 [Ingenia; Philips Healthcare, the Netherlands] or n = 16 [GE HDxt; GE Healthcare]) using an 8-channel head coil. The MRI protocol included axial fast spin-echo T1WI, fast spin-echo T2WI, and post-CET1WI in the axial, coronal, and oblique-sagittal planes. Post-CET1WI were obtained after the intravenous bolus injection of 0.1 mL/kg gadopentetate dimeglumine. Imaging detailed parameters for MRI scan acquisition are shown in Table S1,

Study work flow. (1) MR imaging delineation; (2) extraction of features using the Radcloud radiomics research platform; (3) for feature selection, variance threshold, select KBest method, and LASSO method were used; and (4) development of the radiomics model, clinical model, and visual nomogram. Figure 1 can be viewed online in color at

The Radcloud radiomics research platform (Huiying Medical Technology Co, Ltd, Beijing, China) was used to process imaging data as well as to perform subsequent radiomics statistical analyses.13,21 Tumors were manually delineated in each axial image on the T1WI, T2WI, and CET1WI by a radiologist who had 5 years (radiologist 1) of experience in imaging the head and neck. The regions of interest (ROI) encompassed the entire tumor, including necrotic and cystic changes and hemorrhagic areas, but excluded retinal detachment. The ROIs were manually delineated for the 85 patients who met our inclusion criteria. Volumes of interest (VOIs) were created for each patient by interpolating 2-dimensional ROIs across images encompassing the tumor. The ROIs for each patient were independently reviewed by a senior radiologist with 16 years (radiologist 2) of experience in imaging the head and neck to ensure the accuracy of the ROI to assess the interobserver reproducibility of radiomics feature extraction. The lesion delineation for all the cases was delineated 3 times by 2 radiologists. The interclass correlation coefficient (ICC) of the 1029 features for each MRI sequence was calculated for 85 patients. The ICC ranged from 0.793 to 0.912. Finally, we chose the segmentation with best agreement of features extracted by 2 radiologists as the final lesion area. Because twice segmentation for extracting quantitative features gives similar results, we chose to use 1 segmentation result of a more experienced radiologist for further analysis.13,21

To ensure the stability and reproducibility of the radiomics features, ICC was also calculated for radiomics features between the 2 radiologists for all images. Radiomics features with an ICC of greater than 0.80 were regarded to be in good agreement and retained for further analysis, whereas features with an ICC of less than 0.80 did not undergo further analyses.21 Both radiologists were blinded to patient-specific information.

Feature Extraction

For each MRI sequence, 1029 image features were extracted using a tool from the Radcloud radiomics research platform, which extracted radiomics features from medical image data with a large panel of engineered hard-coded feature algorithms (,21 These image features can be divided into 4 main categories: first-order statistical features, shape features, texture features, and higher-order statistical features. First-order statistics features are related to the intensity distribution in the MRI scan VOI (mean, standard deviation, variance, maximum, median, range, etc). Shape features reflect the shape and size of the VOI (volume, surface area, compactness, maximum diameter, etc). Texture features are related to heterogeneity in intensity distribution. Higher-order statistical features include the first-order statistics and texture features derived from wavelet transformation from the original MRI scans: exponential, square, square root, logarithm, and wavelet (wavelet-LH, wavelet-LHH, wavelet-HLL, wavelet-LLH, wavelet-HLH, wavelet-HHH, wavelet-HHL, and wavelet-LLL). Features are in compliance with the definitions provided by the Imaging Biomarker Standardization Initiative. The details of all feature extraction methods are provided in the Supplementary Data (

Radiomics Feature Selection and Model Building

Feature selection was applied before classification so as to avoid the curse of dimensionality during classification. Feature selection methods included implementing a variance threshold (variance threshold = 0.8) using the SelectKBest method as well as using least absolute shrinkage and selection operator (LASSO) to reduce redundant features. For the variance method, eigenvalues of variance less than 0.8 were removed. In the SelectKBest method, a single-variable-feature selection method, P values were used to evaluate the correlation between the feature and classification results, and only features with P < 0.05 were used. The LASSO method uses a 10-fold cross-validation with a minimum average mean square error to obtain optimal α, the parameter of regularization, which is initialized with an α set to 0.1. Radiomics features with non-zero α coefficients were generated using a training cohort.

The selected features and their relationship with survival were evaluated via univariate Cox analysis. Significant features were weighted with their coefficients, which were weighted using a logistic regression model we noted above and added up to form the rad-score (rad-score = intercept + feature 1 × coefficient 1 + feature 2 × coefficient 2…). Once the median of each patient's rad-score was calculated, patients with higher rad-scores were assigned to the high-risk group and those with lower rad-scores to the low-risk group. The Cox proportional hazards model was used to establish the radiomics model.

Assessment of MRI Findings and Model Building

Magnetic resonance imaging findings of the masses included location, height, largest basal diameter, shape, margin, and signal intensity on T1WI and T2WI compared with gray matter, homogeneity, degree of enhancement within the mass, and secondary retinal detachment. Magnetic resonance imaging features were evaluated by the same radiologists who completed the manual ROI procedures. Magnetic resonance imaging features that were significantly associated with DFS were selected using Cox regression and were built into the clinical model.

Performance of the Radiomics and Clinical Models

Kaplan-Meier survival analysis and the log-rank test were used to evaluate each model's stratification ability. The concordance index (C-index) was used to evaluate each model's discriminative performance.

Radiomics Nomogram Construction and Evaluation

We built a clinical-radiomics nomogram with both radiomics and clinical features. The calibration of the nomogram was assessed using calibration curves and the Hosmer-Lemeshow test. The model's discrimination power was also assessed using the C-index.

Decision curve analysis was used to evaluate the clinical usefulness of the nomogram by calculating the net benefits for a range of threshold probabilities22 using the following formula:

Netbenefit=True positivenFalse positivenpt1pt

where n is the total number of patients in the study and pt is the threshold probability.

Statistical Analysis

Statistical analyses were performed using R 3.3.0 (R Development Core Team, 2016) and SPSS version 17 (SPSS Inc). For clinical characteristics and MRI findings, the independent t test or Mann-Whitney U test was used for continuous variables and the χ2 test or Fisher exact test for categorical variables (2-tailed, P < 0.05). Univariate Kaplan-Meier survival analysis and log-rank test were used to select the most significant clinical variables, where we chose P < 0.1 to be statistically significant so as to avoid missing important clinical information. Multivariate Cox regression analysis using a backward stepwise approach was performed to identify independent risk variables related to the DFS of patients with UM. The X-tile software version 3.6.1 (Yale University School of Medicine, New Haven, CT) was used to choose the optimal cutoff value to stratify patients with UM into a high- or a low-risk group.

Furthermore, the following R packages were used: the psych package was used for calculating ICC; glmnet package for running LASSO-Cox; survival package for building the Cox proportional risk model; ggsurvplot for drawing the Kaplan-Meier curve; rms package for calculating C-index, drawing nomograms and calibration curves; and resourceselection package for the Hosmer-Lemeshow test.


Patient Characteristics

A total of 85 patients with UM were enrolled (Table 1) with a median age of 43 years (range, 18–77 years). The median follow-up time was 60 months (range, 6–125 months). There was no significant difference in clinical characteristics or MRI findings between the training and the validation cohorts (P = 0.10–0.86).

Radiomics Feature Selection and Model Building and Performance

Fifteen radiomics significant features were selected based on MR images using a 10-fold cross-validation LASSO-Cox model: 4 features from T1WI images, 6 features from T2WI images, and 5 features from CET1WI images. The coefficients associated with these significant features are presented in Table 2. The rad-score was calculated for each patient using a weighted linear combination of selected features (rad-score = sum (coefficient (i) × feature (i)) + b (b = 0.237)). Two representative patients with distinctly different DFS times (87 vs 21 months) are shown in Figure 2. Although they had almost the same clinicopathological features, their rad-scores (−0.23 vs 23.46) were significantly different.

TABLE 2 - Radiomics Feature Selection Results Based on the Cross-validation
Categories Individual Features Coefficient
T1-w T1-w-expo_glcm_DifferenceEntropy 1.009
T1-w-log_gldm_LGLEmph 0.776
T1-w-ori_shape_MeshVolume 0.691
T1-w-WHHL_glcm_SumSquares 8.614
T2-w T2-w-expo_glcm_ClusterTendency 0.075
T2-w-expo_glcm_Contrast 1.757
T2-w-square_fos_Energy 2.668
T2-w-WHLH_glszm_LargeAreaEmph 1.098
T2-w-WHLL_ngtdm_Strength 2.052
T2-w-WLLL_gldm_SDHGLEmph 0.784
CET1-w CET1-w-sq_glszm_SZNonUni 1.647
CET1-w-WHHH_glrlm_GLNonUni 1.960
CET1-w-WHHL_fos_90P 0.379
CET1-w-WHLH_glrlm_ShortRHGLEmph 2.455
CET1-w-WHLL_glrlm_ShortRHGLEmph 2.198
glcm, gray level co-occurrence matrix; gldm, gray level dependence matrix; glrlm, gray level run length matrix; ngtdm, neighboring gray tone difference matrix; glszm, gray level size zone matrix; T1-w-exponential_glcm_DifferenceEntropy, T1-w-expo_glcm_DifferenceEntropy; T1-w-logarithm_gldm_LowGrayLevelEmphasis, T1-w-log_gldm_LGLEmph; T1-w-original_shape_MeshVolume; T1-w-ori_shape_MeshVolume; T1-w-wavelet.HHL_glcm_SumSquares, T1-w-WHHL_glcm_SumSquares; T2-w-exponential_glcm_ClusterTendency, T2-w-expo_glcm_ClusterTendency; T2-w-exponential_glcm_Contrast; T2-w-expo_glcm_Contrast; T2-w-square_firstorder_Energy, T2-w-square_fos_Energy; T2-w-wavelet.HLH_glszm_LargeAreaEmphasis, T2-w-wavelet.HLL_ngtdm_Strength, T2-w-WHLL_ngtdm_Strength; T2-w-wavelet.LLL_gldm_SmallDependenceHighGrayLevelEmphasis, T2-w-WLLL_gldm_SDHGLEmph; CET1-w-square_glszm_SizeZoneNonUniformity, CET1-w-sq_glszm_SZNonUni; CET1-w-wavelet.HHH_glrlm_GrayLevelNonUniformity, CET1-w-WHHH_glrlm_GLNonUni; CET1-w-wavelet.HHL_firstorder_90Percentile, CET1-w-WHHL_fos_90P; CET1-w-wavelet.HLH_glrlm_ShortRunHighGrayLevelEmphasis, CET1-w-WHLH_glrlm_ShortRHGLEmph; CET1-w-wavelet.HLL_glrlm_ShortRunHighGrayLevelEmphasis, CET1-w-WHLL_glrlm_ShortRHGLEmph.
“HHL,” “HLH,” “HLL,” “LLL,” and “HHH” represent high-pass filter and low-pass filter on the X, Y, and Z 3-dimensionally. “H” represents high-pass filter, and “L” represents low-pass filter. X, Y, and Z directions are relative to the standard DICOM LPS (left-posterior-superior) coordinate system.

Patient 1. A 42-year-old woman with UM in the right globe. A well-defined mound mass shows hyperintensity on T1WI (A) and heterogeneous enhancement on CET1WI (B). The height and basal diameter of the mass were 8 and 13 mm, respectively. The DFS of the patient was 87 months. The rad-score was −0.232. Patient 2. A 32-year-old woman with UM in the left globe. A well-defined mound mass hyperintensity on T1WI (C) and heterogeneous enhancement on CET1WI (D). The height and basal diameter of the mass were 10 and 13 mm, respectively. The DFS of the patient was 21 months. The rad-score was 23.464. Figure 2 can be viewed online in color at

The median hazard score of the training cohort was used to separate patients into a high- or a low-risk group (median of rad-score = 1.06). The Kaplan-Meier survival curves of the training and validation cohorts featured good stratification (log-rank test, P = 0.009 and 0.02, respectively; Fig. 3A). The radiomics model performed well in the training cohort, yielding a C-index of 0.730 (95% confidence interval [CI], 0.604–0.816), and demonstrated efficacy in predicting the risk of DFS in the validation cohort (C-index, 0.640; 95% CI, 0.517–0.763).

Kaplan-Meier analysis of the radiomics model (A) and clinical model (B) in the training cohort (left) and validation cohort (right). Vertical lines indicate censored data. Shadows represent 95% CI. P values were calculated using the log-rank test. Figure 3 can be viewed online in color at

Cox Regression Models of Clinical Characteristics and MRI Findings

Univariate analysis revealed age, largest basal diameter, height, homogeneity of enhancement, and retinal detachment to be significant biomarkers (P < 0.001, P = 0.027, P = 0.033, P = 0.069, P = 0.011, respectively) (Table S2, Multivariate Cox regression analysis verified that only age (P = 0.002; hazard ratio [HR], 4.886), height (P = 0.036; HR, 2.487), and largest basal diameter (P = 0.028; HR, 0.305) were independent risk factors related to the DFS of patients with UM. The optimum cutoff values of age, basal diameter, and height of patients with UM generated by the X-tile plot were 43 years, 11 mm, and 10 mm, respectively. Further, these factors were incorporated to build a clinical prediction model, yielding a C-index of 0.691 (95% CI, 0.571–0.811) in the training cohort and 0.826 (95% CI, 0.673–0.978) in the validation cohort. This model was able to stratify high- and low-risk DFS in the validation cohort (P = 0.009); however, the stratification was poor in the training cohort (P = 0.036, Fig. 3B).

Construction of the Radiomics Nomogram and Validation

We constructed a clinical-radiomics nomogram that integrated age, basal diameter, height, and rad-score to predict the DFS of UM (Fig. 4A). The calibration curve of the nomogram demonstrated good agreement between prediction and observation in the training cohort (Fig. 4B). The Hosmer-Lemeshow test yielded a nonsignificant statistic (P = 0.927), suggesting that there was no departure from perfect fit. The C-index for the prediction nomogram was 0.741 (95% CI, 0.637–0.845) in the training cohort and 0.912 (95% CI, 0.847–0.977) in the validation cohort.

The nomogram (A) was built to predict the DFS of patients with UM. Note: The 1–2 age is defined as follows: 1, <43 years, and 2, ≥43 years; the calibration curves (B) of the nomogram demonstrated the predictive performance of survival in UM.

Clinical Use

Decision curve analysis was used to evaluate the benefit of the clinical, radiomics, and clinical-radiomics nomograms. For a threshold probability of less than 58% (ie, if the survival rate of patients with UM was <58%), we found that patients would benefit more from the radiomics nomogram than from either clinical decision (ie, MRI findings or follow-up) or no clinical intervention (Fig. 5).

Decision curve analysis for the radiomics and clinical nomograms. The dark gray line describes the scheme of no clinical intervention. The light gray line illustrates the scheme of clinical intervention. The red, blue, and orange lines represent clinical, radiomics, and clinical-radiomics nomograms, respectively. The x axis is the threshold probability, whereas the y axis is the net benefit. Decision curve analysis revealed that if the threshold probability was greater than 10%, using the clinical-radiomics combined prediction model was more beneficial than using the clinical model. Figure 5 can be viewed online in color at


In the present study, we evaluated the prognostic value of pretreatment MRI-based radiomics features of patients with UM. The results demonstrated that radiomics signature had a good predictive performance in DFS estimation with a C-index of 0.730 in the training cohort, thereby allowing the classification of patients into either a high- or low-risk group (P < 0.05). A nomogram combining the rad-score and visible MRI findings was used to evaluate patients on an individual basis and also provided visual feedback to estimate the DFS of each patient with UM. Decision curve analysis was also performed to confirm the clinical benefit, which demonstrated that the radiomics nomogram was superior to the clinical model.

The radiomics model comprised 15 significant features selected from the high-throughput textures of the 3-dimensional volume of the tumor, which provided prognostic information for evaluating risk stratification for tumor progression. The radiomics features provided a general condition of intratumor heterogeneity and the surrounding microenvironment.23,24 The significant features are associated with the biological and genetic characteristics of the tumor, allowing prognostic prediction.25 In the current study, we screened 15 radiomics features, including T1-w-ori_shape_MeshVolume, T2-w-expo_glcm_Contrast, T2-w-square_fos_Energy, CET1-w-WHHL_fos_90P, and CET1-w-WHHH_glrlm_GLNonUni. The shape feature, MeshVolume, reflected the volume of ROI; in other words, it presented the volume of the tumor. As reported in literature, a large tumor side is linked with worse patient outcome.26–28 The “Energy” feature was highly consistent with normal radiological experience, which describes the external contour characteristics of the tumor.29 High intensity of tumor enhancement was linked to the invasiveness of the tumor and in turn with poor patient prognosis. Furthermore, low GLNonUni values were associated with patients with UM with short DFS. The present study demonstrated that the gray-level nonuniformity feature was found to be significantly predictive of prognosis: lower uniformity of gray levels indicated worse prognoses, high tumor malignancy, and more tumor metastasis, in line with previous studies.30,31 In addition, GLCM and GLRLM, which were calculated using spatial gray-level dependence matrices in tumor ROI, could quantitatively respond to tumor heterogeneity and uniformity.32 The radiomics signature based on rad-score in our study achieved a good predictive value, as demonstrated by a C-index of 0.640 in the validation cohort. The Kaplan-Meier curve suggested that the 5-year DFS was 91.5% for the low-rad-score group and 56.3% for the high-rad-score group, with good risk stratification. The MRI radiomics model can predict the survival of patients with gastric cancer, nasopharyngeal carcinoma, and cervical cancer and pretreatment in many studies.21,33,34 To the best of our knowledge, this is the first study to report that radiomics features based on MRI acquired before surgical treatment can predict the survival of preoperative patients with UM.

In the present study, we also evaluated the value of MRI findings for the prediction of the DFS of patients with UM. To date, the correlation between MRI findings and survival time in patients with UM has been seldom studied and remains unknown. Our study found that height (P = 0.036; HR, 2.487) and largest basal diameter (P = 0.028; HR, 0.305) were independent risk factors related to UM progression. The threshold values for the tumor's largest basal diameter of 11 mm and the height of 10 mm indicated that there are 2 UM classes that differ with respect to the risk of survival outcomes. Basal diameter and height were the 2 risk factors that correlated with UM progression, which are still important and routinely used in clinical settings. Kim et al35 reported that a tumor basal diameter of 15 mm or greater and a vertical depth of 10 mm or greater, based on ultrasound, were associated with the distant recurrence of UM. However, it is anticipated that MRI will provide better soft-tissue contrast and clearly resolve the tumor compared with ultrasound because ultrasound depends on the skill of the operator; therefore, MRI measurement of diameter was more accurate than that of ultrasound.36 We further incorporated radiomics signature into MRI findings to construct a visual clinical-radiomics prediction model, which displayed better prediction performance, with a C-index of 0.741. These results suggest that the radiomics signature reinforces the prognostic ability of clinical diagnosis. Ultimately, patients with the same MRI findings might be classified into different risk groups on the basis of the radiomics model, and thus, a different clinical management plan might be explored to improve their survival outcomes. As a result, the noninvasive radiomics signature, which made use of the MRI already available, could serve as a more convenient biomarker for the prediction of the DFS of patients with UM.

Our data also indicated that a nomogram combining rad-score, age, basal diameter, and height serves as a visual tool that aids clinicians in making clinical decisions. The results demonstrated good calibration and discrimination in both the training and validation cohorts. The comparable discrimination implied that the nomogram was robust in quantifying an individual's risk of poor outcome. Using the clinical-radiomics nomogram, an estimated probability of patient survival in UM could be calculated after referring to the MRI rad-scores as well as other clinical information. The current study is the first to compare the predictive performance of a radiomics nomogram and MRI features in UM. Notably, a valuable feature of the clinical-radiomics nomogram was its discriminatory ability in patients with UM with small tumor masses. Patients with UM diagnosed with a small tumor were typically considered to be at a low risk of poor outcome. A tumor size of 1 to 2.5 mm as the apical height and 5 to 16 mm as the largest basal diameter is the definition for small UM according to the Collaborative Ocular Melanoma Study.37 However, some patients with small UM also present with liver metastasis.28 It was a formidable challenge to precisely identify which patients will experience liver metastasis because of the short time from detection of metastasis to death. Encouragingly, our nomogram showed good discriminatory ability for patients with UM with small tumors. For instance, a patient diagnosed with a small UM, when categorized into the high-risk group based on the cutoff values of the risk score derived from the nomogram, subsequently had a significantly greater probability of worse outcome. The ability to predict patients' survival can help in determining whether more intensive observation and aggressive treatment regimens should be administered, which might improve the clinical outcome of patients with UM. Therefore, our clinical-radiomics nomogram may serve as an accurate, preoperatively noninvasive, and reliable predictive tool for UM progression, particularly for patients with small UM.

This study has several limitations. First, the study contained a small number of enrolled patients; therefore, despite our efforts to prevent overfitting and generalizability (features standardization and 10-fold cross-validation in LASSO), the prediction model does not have a good prediction effect in the validation cohort. Therefore, the model needs to be further verified with larger data sets. Accordingly, multicenter clinical trials with larger sample size are needed to improve the predictive power of our model. Second, various MRI scanners were used for the extraction of radiomics features. We noted that the use of different field strengths was beneficial in terms of applicability to other sites and centers33; furthermore, the underlying biology of tumor tissue will not change owing to variations in acquisition and image reconstruction parameters, as reported by Gillies et al.23 Third, our study included only patients who underwent enucleation or resection and did not include those who underwent other treatments, such as plaque radiotherapy, which is the most common treatment for patients with UM. Therefore, whether the prediction model is suitable for patients with UM who receive other therapeutic treatments needs to be further verified. Lastly, our study focused on a limited number of clinical features. In the future, when tumor information is routinely collected, such as genetic information and other molecular tumor markers, the use of these types of predictive models will become increasingly important and will improve the clinical efficacy of the model for patients with UM.

In conclusion, we have established a radiomics signature based on MRI to predict the risk of DFS in pretreatment UM, which may serve as a useful and noninvasive tool for clinicians to identify patients at a high risk of early progression and may be valuable to the patient and family to permit personal planning and arrangements for future medical care. In addition, the radiomics nomogram may serve as a potential tool to guide individual care.


1. Chattopadhyay C, Kim DW, Gombos DS, et al. Uveal melanoma: from diagnosis to treatment and the science in between. Cancer. 2016;122:2299–2312.
2. Eagle RC Jr. The pathology of ocular cancer. Eye (Lond). 2013;27:128–136.
3. Diener-West M, Reynolds SM, Agugliaro DJ, et al. Screening for metastasis from choroidal melanoma: the Collaborative Ocular Melanoma Study Group Report 23. J Clin Oncol. 2004;22:2438–2444.
4. Everett L, Copperman T. Metastatic uveal melanoma. N Engl J Med. 2019;380:1853.
5. Singh AD, Turell ME, Topham AK. Uveal melanoma: trends in incidence, treatment, and survival. Ophthalmology. 2011;118:1881–1885.
6. Shields CL, Furuta M, Thangappan A, et al. Metastasis of uveal melanoma millimeter-by-millimeter in 8033 consecutive eyes. Arch Ophthalmol. 2009;127:989–998.
7. Robertson AG, Shih J, Yau C, et al. Integrative analysis identifies four molecular and clinical subsets in uveal melanoma. Cancer Cell. 2018;33:151.
8. Field MG, Decatur CL, Kurtenbach S, et al. PRAME as an independent biomarker for metastasis in uveal melanoma. Clin Cancer Res. 2016;22:1234–1242.
9. Smit KN, Jager MJ, de Klein A, et al. Uveal melanoma: towards a molecular understanding. Prog Retin Eye Res. 2020;75:100800.
10. Kamrava M, Sepahdari AR, Leu K, et al. Quantitative multiparametric MRI in uveal melanoma: increased tumor permeability may predict monosomy 3. Neuroradiology. 2015;57:833–840.
11. Foti PV, Longo A, Reibaldi M, et al. Uveal melanoma: quantitative evaluation of diffusion-weighted MR imaging in the response assessment after proton-beam therapy, long-term follow-up. Radiol Med. 2017;122:131–139.
12. Lambin P, Rios-Velazquez E, Leijenaar R, et al. Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer. 2012;48:441–446.
13. Mao J, Fang J, Duan X, et al. Predictive value of pretreatment MRI texture analysis in patients with primary nasopharyngeal carcinoma. Eur Radiol. 2019;29:4105–4113.
14. Giesel FL, Schneider F, Kratochwil C, et al. Correlation between SUVmax and CT radiomic analysis using lymph node density in PET/CT-based lymph node staging. J Nucl Med. 2017;58:282–287.
15. Liu Z, Li Z, Qu J, et al. Radiomics of multiparametric MRI for pretreatment prediction of pathologic complete response to neoadjuvant chemotherapy in breast cancer: a multicenter study. Clin Cancer Res. 2019;25:3538–3547.
16. Giannini V, Mazzetti S, Bertotto I, et al. Predicting locally advanced rectal cancer response to neoadjuvant therapy with 18F-FDG PET and MRI radiomics features. Eur J Nucl Med Mol Imaging. 2019;46:878–888.
17. Bi WL, Hosny A, Schabath MB, et al. Artificial intelligence in cancer imaging: clinical challenges and applications. CA Cancer J Clin. 2019;69:127–157.
18. Su Y, Xu X, Zuo P, et al. Value of MR-based radiomics in differentiating uveal melanoma from other intraocular masses in adults. Eur J Radiol. 2020;131:109268.
19. Fry MV, Augsburger JJ, Corrêa ZM. Clinical features, metastasis, and survival in patients younger than 21 years with posterior uveal melanoma. JAMA Ophthalmol. 2019;137:75–81.
20. Amin MB, Greene FL, Edge SB, et al. The eighth edition AJCC Cancer Staging Manual: continuing to build a bridge from a population-based to a more “personalized” approach to cancer staging. CA Cancer J Clin. 2017;67:93–99.
21. Shu J, Tang Y, Cui J, et al. Clear cell renal cell carcinoma: CT-based radiomics features for the prediction of Fuhrman grade. Eur J Radiol. 2018;109:8–12.
22. Dong D, Tang L, Li ZY, et al. Development and validation of an individualized nomogram to identify occult peritoneal metastasis in patients with advanced gastric cancer. Ann Oncol. 2019;30:431–438.
23. Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, they are data. Radiology. 2016;278:563–577.
24. Limkin EJ, Sun R, Dercle L, et al. Promises and challenges for the implementation of computational medical imaging (radiomics) in oncology. Ann Oncol. 2017;28:1191–1206.
25. Song J, Shi J, Dong D, et al. A new approach to predict progression-free survival in stage IV EGFR-mutant NSCLC patients with EGFR-TKI therapy. Clin Cancer Res. 2018;24:3583–3592.
26. Rietschel P, Panageas KS, Hanlon C, et al. Variates of survival in metastatic uveal melanoma. J Clin Oncol. 2005;23:8076–8080.
27. Nathan P, Cohen V, Coupland S, et al. Uveal melanoma UK national guidelines. Eur J Cancer. 2015;51:2404–2412.
28. Shields CL, Kaliki S, Furuta M, et al. Diffuse versus nondiffuse small (≤ 3 mm thickness) choroidal melanoma: comparative analysis in 1,751 cases. The 2012 F. Phinizy Calhoun lecture. Retina. 2013;33:1763–1776.
29. Shen C, Liu Z, Wang Z, et al. Building CT radiomics based nomogram for preoperative esophageal cancer patients lymph node metastasis prediction. Transl Oncol. 2018;11:815–824.
30. van Dijk LV, Thor M, Steenbakkers R, et al. Parotid gland fat related magnetic resonance image biomarkers improve prediction of late radiation-induced xerostomia. Radiother Oncol. 2018;128:459–466.
31. Mo X, Wu X, Dong D, et al. Prognostic value of the radiomics-based model in progression-free survival of hypopharyngeal cancer treated with chemoradiation. Eur Radiol. 2020;30:833–843.
32. Liu Z, Zhang XY, Shi YJ, et al. Radiomics analysis for evaluation of pathological complete response to neoadjuvant chemoradiotherapy in locally advanced rectal cancer. Clin Cancer Res. 2017;23:7253–7262.
33. Zhang L, Dong D, Li H, et al. Development and validation of a magnetic resonance imaging–based model for the prediction of distant metastasis before initial treatment of nasopharyngeal carcinoma: a retrospective cohort study. EBioMedicine. 2019;40:327–335.
34. Fang J, Zhang B, Wang S, et al. Association of MRI-derived radiomic biomarker with disease-free survival in patients with early-stage cervical cancer. Theranostics. 2020;10:2284–2292.
35. Kim JH, Shin SJ, Heo SJ, et al. Prognoses and clinical outcomes of primary and recurrent uveal melanoma. Cancer Res Treat. 2018;50:1238–1251.
36. Foti PV, Travali M, Farina R, et al. Diagnostic methods and therapeutic options of uveal melanoma with emphasis on MR imaging—part II: treatment indications and complications. Insights Imaging. 2021;12:67.
37. Design and methods of a clinical trial for a rare condition: the Collaborative Ocular Melanoma Study. COMS report no. 3. Control Clin Trials. 1993;14:362–391.

magnetic resonance imaging; prognosis; radiomics; uveal melanoma

Supplemental Digital Content

Copyright © 2022 Wolters Kluwer Health, Inc. All rights reserved.