Correlation Analysis of Apparent Diffusion Coefficient Histogram Parameters and Clinicopathologic Features for Prognosis Prediction in Uveal Melanoma

Abstract

Purpose: To investigate the correlation between apparent diffusion coefficient (ADC) histograms and high-risk clinicopathologic features related to uveal melanoma (UM) prognosis.

Methods: This retrospective study included 53 patients with UM who underwent diffusion-weighted imaging (DWI) between August 2015 and March 2024. Axial DWI was performed with a single-shot spin-echo echo-planar imaging sequence. ADC histogram parameters of ADC_mean, ADC_50%, interquartile range (IQR), skewness, kurtosis, and entropy were obtained from DWI. The relationships between histogram parameters and high-risk clinicopathological characteristics including tumor size, preoperative retinal detachment, histological subtypes, Ki-67 index, and chromosome status, were analyzed by Spearman correlation analysis, Mann–Whitney U test, or Kruskal–Wallis test.

Results: A total of 53 patients (mean ± SD age, 55 ± 15 years; 22 men) were evaluated. The largest basal diameter (LBD) was correlated with kurtosis (r = 0.311, P = 0.024). Tumor prominence (TP) was correlated with entropy (r = 0.581, P < 0.001) and kurtosis (r = 0.273, P = 0.048). Additionally, significant correlations were identified between the Ki-67 index and ADC_mean (r = −0.444, P = 0.005), ADC_50% (r = −0.487, P = 0.002), and skewness (r = 0.394, P = 0.014). Finally, entropy was correlated with monosomy 3 (r = 0.541, P = 0.017).

Conclusions: The ADC histograms provided valuable insights into high-risk clinicopathologic features of UM and hold promise in the early prediction of UM prognosis.

Uveal melanoma (UM) is the most common and highly malignant primary tumor of the adult eye.¹ UM has an insidious onset and is generally difficult to detect in the early stages, and metastasis occurs in approximately 50% of patients as the tumor progresses.² More than 90% of patients die within 2 years after metastasis occurs.³ Considering the challenging characteristics of UM, early diagnosis and prognostic evaluation of these extremely metastatic tumors are crucial. 

In contrast to most tumors, UM is usually not diagnosed by biopsy, as clinical presentation and imaging studies are usually employed to establishing a UM diagnosis.^4,5 Magnetic resonance imaging (MRI) examination is an important imaging and therapeutic decision-making basis for UM due to its high soft tissue contrast, spatial imaging capability, and ability to generate functional images.^6–10 Among the various MRI scanning sequences, diffusion-weighted imaging (DWI), which is based on the Brownian motion of water molecules, is crucial for tumor diagnosis in many aspects.^11–14 DWI presents quantitative information as an apparent diffusion coefficient (ADC).^13,15 The ADC helps differentiate ocular melanoma from retinal detachment.¹² A previous study has reported that early changes in ADC values are significantly correlated with good tumor response to proton-beam therapy.¹⁶ Additionally, ADC serves as an independent risk factor for early postoperative recurrence of hepatocellular cancer.¹⁷ These findings suggest that DWI sequences can provide additional information. However, single ADC values cannot reflect the heterogeneity of the entire tumor and are influenced by the scanning machine and parameters.¹⁸ 

Recently, DWI-derived ADC histograms have been receiving increasing attention in tumor research.^19–21 The ADC histogram and the characteristic parameters related to the distribution of ADC values are obtained by acquiring and calculating the ADC values of whole-tumor volume on DWI images. They reflect the internal characteristics of the tumor, which cannot be assessed using a single ADC value.¹⁸ Moreover, ADC histograms and histogram-derived parameters correlate with the clinical and pathological features of tumors, provide a predictive value for prognosis, and potentially predict response to treatment. For example, ADC histogram parameters can predict meningioma grade, subtype, and proliferative activity.²² Further, preoperative ADC histogram analysis has shown potential as a reliable imaging-based method for prognostic prediction in patients with glioblastoma.²³ These results are encouraging and highlight the clinical significance of ADC histogram as a noninvasive imaging test for tumors. However, ADC histogram analysis has not yet been applied to predict UM prognosis. 

Based on the above inspiring findings of ADC histograms and the challenges faced in UM clinical practice, we propose to provide insights for UM prognostic assessment by using ADC histogram analysis for early identification of high-risk clinical and histopathologic features. In this study, we first extracted ADC histograms and histogram parameters from DWI images. Subsequently, we investigated the utility of ADC histogram in high-risk clinicopathological factors, including tumor size, preoperative retinal detachment (RD), histological subtypes, Ki-67 index, and chromosome status, which can help to predict UM prognosis. 

Materials and Methods

Patients

This retrospective study was approved by the Medical Ethics Committee of Shanghai Ninth People's Hospital (SH9H-2023-T462-1). Between August 2015 and March 2024, 72 patients with UM were initially identified at our center based on the following inclusion criteria: (1) clinical or pathologic diagnosis of UM, and (2) MRI scans received at our center. Patients were then screened according to the following exclusion criteria: (1) MRI scans did not contain DWI sequences (n = 14), (2) MRI images had obvious artifacts and inadequate image quality (n = 4), and (3) clinical treatment occurred before orbital MRI (n = 1). The final number of patients included in the study was 53. Figure 1 shows the process of patient inclusion and exclusion. 

Figure 1.

 

Flow chart for the inclusion and exclusion of participants.

MRI Scanning Protocol

The patient was placed supine, and MRI was performed on a 3T MRI scanner (MAGNETOM Verio; Siemens, Munich, Germany) using with a 64-channel head and neck coil. To minimize motion artifacts, the patient's head was immobilized on both sides with pillows, and the patient was instructed to fixate his or her vision on a central point during image acquisition. The orbital MRI scanning sequences included conventional plain scanning, dynamic enhancement, and DWI sequences. The main sequences and parameters were as follows: (1) transverse T1-weighted imaging (T1WI) with a fast spin-echo sequence, repetition time (TR) of 620 ms, echo time (TE) of 9.2 ms, slice thickness of 3 mm, and field of view (FOV) of 100 mm × 100 mm; (2) transverse T2–short tau inversion recovery (T2-STIR) sequence, TR of 4000 ms, and TE of 75 ms; (3) axial DWI using single-shot spin-echo echo-planar imaging (EPI) sequence, with a TR of 3500 ms, TE of 65 ms, flip angle of 90°, FOV of 180 mm × 180 mm, slice thickness of 3 mm, interslice gap of 0.5 mm, b values of 0 and 1000 s/mm², matrix of 256 × 320, scan time of 150 seconds, and gradient directions of x, y, and z, as well as an ADC map calculated with a mono-exponential fit; and (4) dynamic enhancement with a breath-hold sequence using a volumetric interpolator, along with a TR of 5.0 ms, TE of 1.97 ms, slice thickness of 2 mm, and scanning time of 4 minutes. Additionally, eight time phases were acquired. Furthermore, enhanced transverse, coronal, and oblique sagittal T1WI scans were performed with a TR of 433 ms and a TE of 9.2 ms. A gadolinium contrast agent (dose, 14 mL) was injected intravenously at a rate of 2.5 mL/s using a high-pressure syringe. All patients included in the study had been scanned with the above MRI scanning protocol since 2015. 

ADC Histogram Feature Extraction

All images were transferred to a Siemens workstation, where T1WI and DWI images and ADC maps were exported from the picture archiving and communication system (PACS) server in the Digital Imaging and Communications in Medicine (DICOM) format. Image analysis was conducted separately by two imaging radiologists who had >5 years of experience in diagnosing ocular diseases. Finally, all analyses were checked by a radiologist with >10 years of imaging experience. They were blinded to the patients’ clinical and pathologic information. 

The images in DICOM format were first imported into FireVoxel software (https://firevoxel.org/download/) and normalized using N4 algorithms to remove nonuniformity before analyzing the images. For patients whose ADC maps were automatically generated at the time of DWI scanning, regions of interest (ROIs) were outlined layer by layer on the ADC maps referring to the lesions on T1WI, followed by generation of three-dimensional (3D) ROIs by the software. ROIs were delineated on the DWI images for patients without ADC maps, after which ADC maps of the ROIs were generated by the FireVoxel software. Voxels containing peritumor tissue, such as vitreous, sclera, or tumor margin, were discarded when the ROIs were drawn. Retinal detachment lesions were also avoided. The number of voxels within the ROIs and the ROI volume for each tumor are shown in Supplementary Table S1. Subsequently, FireVoxel software was employed to construct ADC histograms and calculate the following histogram-derived parameters: ADC mean (ADC_mean), ADC 50th percentile (ADC_50%), interquartile range (IQR), skewness, kurtosis, and entropy. 

Skewness represents the distribution of the histogram, in which normal distribution is considered to have a skewness of 0.²⁴ Skewness > 0 indicates that the histogram is positively skewed (i.e., the right tail is extended). Conversely, skewness < 0 indicates that the histogram has a longer left tail (i.e., negatively skewed). Skewness was calculated using:  

\begin{equation*} {\rm{Skewness\ = \ }}\frac{{\frac{1}{n}\mathop \sum \nolimits_{i = 1}^n {{{\left( {{{x}_i} - \bar{x}} \right)}}^3}}}{{{{{\rm{\sigma }}}^2}}} \end{equation*}

 

Kurtosis conveys the features of the histogram peak, with kurtosis > 0 indicating steeper than normal distribution and kurtosis < 0 indicating smoother than normal distribution:²⁴ 

\begin{equation*} {\rm{Kurtosis}} = \frac{{\frac{1}{n}\mathop \sum \nolimits_{i = 1}^n {{{\left( {{{x}_i} - \bar{x}} \right)}}^4}}}{{{{{\rm{\sigma }}}^4}}} - 3 \end{equation*}

 

Entropy indicates irregularities in the distribution of voxels within a tumor and allows for quantifying heterogeneity within a tumor:²⁵ 

\begin{equation*} {\rm{Entropy}} = - \mathop \sum \limits_{i = 1}^n p\left( {{{x}_i}} \right)\log p\left( {{{x}_i}} \right) \end{equation*}

 

In these equations, n is the total number of voxels, the parameter x_i is the ADC value of individual voxels, \(\bar{x}\) represents the mean ADC value, σ represents the standard deviation, and p(x_i) represents the probability that the ADC value is x_i. 

Clinical/Histological Data Collection

Ocular ultrasound reports were reviewed retrospectively for reported largest base diameter (LBD) and tumor prominence (TP) of the tumor on ultrasound imaging. For ultrasound tests that were performed at other centers with no recording of tumor size, tumor sizes were extracted from preoperative MRI reports. Tumor stage was further assessed according to LBD and TP. The T stage of tumors was determined based on the American Joint Committee on Cancer (AJCC), Eighth Edition. Tumors were classified as small, medium, or large based on the Collaborative Ocular Melanoma Study (COMS) stage. Preoperative RD was extracted from the electronic history. 

As for histopathological information, including histological subtypes, Ki-67 index, monosomy 3, and 8q amplification, were assessed by two pathologists with >5 years of experience. Tissue blocks containing tumor were obtained after surgical resection or enucleation and then sectioned at thicknesses of 5 µm. Histological subtype was assessed by hematoxylin and eosin (H&E) staining. The expression of Ki-67, which is an important pathological indicator that reflects the proliferative activity of tumor cells, was estimated by immunohistochemical staining. Monosomy 3 and 8q amplification were obtained by fluorescence in situ hybridization. 

Statistical Analysis

All statistical analyses were performed using SPSS Statistics 26 (IBM, Chicago, IL, USA) and Prism 8 (GraphPad, Boston, MA, USA). The normality and homoscedasticity of the data were tested using the Kolmogorov–Smirnov method and Levene's test, respectively. Between-group comparisons of the ADC parameters were conducted via the Mann–Whitney U test and Kruskal–Wallis test. Correlation coefficients and P values were obtained utilizing Spearman correlation analysis, with the correlation coefficient being denoted by the letter r. A two-tailed P < 0.05 was considered statistically significant. All statistical analyses were confirmed by Hui Wang, an expert in statistics. 

Results

Patient Characteristics

A total of 53 patients were included in the analysis, including 22 males and 31 females. The mean age of the patients was 55 years. There were more patients with left eye involvement than right eye involvement (29 vs. 24). Other characteristics such as tumor size and stage, are summarized in Table 1. 

Table 1.

 

Demographic Information for the Participants with UM

 

ADC Histogram Parameters

Using ROIs drawn on DWI images, we obtained ADC histograms and their derived parameters for each tumor. Figure 2 illustrates the ROIs and ADC histograms for three patients with UM. Table 2 lists the ADC histogram parameters of all included patients. 

Figure 2.

 

MRI images and ADC histograms of three patients with UM. (A) The orbital MRI images and whole-volume ADC histogram of a 76-year-old male with UM of the left eye. (B) The orbital MRI images and whole-volume ADC histogram of a 71-year-old female with UM of the right eye. (C) The orbital MRI images and whole-volume ADC histogram of a 56-year-old male with UM of the right eye. The white arrows indicate locations of the tumors, and the red areas represent the ROIs.

Table 2.

 

ADC Histogram Features of All Patients With UM

 

ADC Histogram Parameters and Clinical Characteristics

Tumor size is a prominent clinical feature of UM that has been suggested to correlate with prognosis and help inform the choice of treatment modality. Accordingly, we first analyzed the correlation of ADC parameters with the LBD and TP. Our results indicated that entropy (r = 0.581, P < 0.001) and kurtosis (r = 0.273, P = 0.048) were significantly correlated with TP, whereas kurtosis was significantly correlated with LBD (r = 0.311, P = 0.024) (Table 3). Scatterplots for all correlations are shown in Supplementary Figure S1. 

Table 3.

 

Association Between ADC Histogram Parameters and Clinical Characteristics

 

Preoperative RD is a common UM complication associated with poor visual outcomes. Preoperative RD was detected in 23 of our patients (43%). Further, we analyzed the correlation of preoperative RD with ADC histogram parameters (Table 3) and found no significant difference in these parameters between patients with preoperative RD and those without preoperative RD. 

ADC Histogram Parameters and Pathological Characteristics

In terms of histologic subtypes, UM was classified as epithelioid cell type (31%), spindle cell type (46%), and mixed cell type (23%). We further explored the relationship between ADC histogram parameters and UM histologic subtypes. The results showed no statistical differences between histogram parameters and histologic subtypes (Supplementary Table S2). 

Subsequently, we analyzed the relationship between the Ki-67 index and ADC parameters. Our results showed that ADC_mean (r = −0.444, P = 0.005), ADC_50% (r = −0.487, P = 0.002), and skewness (r = 0.394, P = 0.014) were significantly correlated with the Ki-67 index. Scatterplots of the correlations between Ki-67 and histogram parameters are shown in Supplementary Figure S1. 

Common chromosomal abnormalities associated with UM include monosomy 3 and 8q amplification. Between these, monosomy 3 status is considered to be an extremely important prognostic marker linked to poor prognosis of UM.²⁶ Therefore, we further examined whether the ADC histogram parameters could also act as markers of chromosome 3 or 8 abnormalities. Our correlation analysis suggested that entropy was positively correlated with the percentage of monosomy 3 (r = 0.541, P = 0.017) but not with chromosome 8q amplification. Scatterplots of the correlations between monosomy 3 and histogram parameters are shown in Supplementary Figure S1. 

Discussion

In this study, we found that the ADC histogram parameters were significantly correlated with UM size, Ki-67 index, and monosomy 3 status, indicating their value for assessing the survival prognosis of UM. Conversely, the ADC histogram parameters did not correlate with preoperative RD, which is related to visual prognosis. In current clinical practice, the ROI is selected at the individual image level, with the average ADC value within the ROI representing the ADC value for the entire tumor. However, the ADC value depends on scanning parameters; thus, differences in MRI scanners or ROIs may result in modified ADC values. Furthermore, ROI selection may not adequately reflect the heterogeneity within the tumor. These factors may explain the contradictory results of the various studies using ADC values from different cohorts to study the same tumor type.^27,28 In contrast, the ADC histogram, which standardizes the image and reflects the distribution of ADC values within the entire lesion, is a more accurate and reliable indicator. 

ADC histograms have promising applications in other tumors, and multiple studies have found that ADC histogram parameters can predict prognosis and treatment response.^23,29 These findings have great clinical significance, because ADC histograms serve as clinically accessible non-invasive imaging markers that can provide comprehensive guidance for diagnostic and therapeutic decisions. For UM, the diagnosis depends heavily on imaging, as it is not usually confirmed by biopsy. Moreover, the anatomical location of UM in the posterior eye region limits the application of some existing adjuvant investigations. As a result, the diagnosis and evaluation of UM depend heavily on MRI. In addition, MRI is essential for selecting treatment modalities, developing radiation treatment plans, and providing post-treatment follow-up for UM.⁷ Therefore, improving the diagnostic value of MRI is critical and urgent for the early diagnosis and risk classification of UM, ultimately enhancing treatment outcomes. In this study, we obtained MRI-derived ADC histogram parameters, including ADC_mean, ADC_50%, IQR, skewness, kurtosis, and entropy, and we evaluated the correlation of these parameters with high-risk clinicopathologic features related to UM prognosis. 

UM size is the most crucial feature in the clinic setting because LBD and TP are associated with poor prognosis.³⁰ Additionally, UM treatment is guided by UM size, particularly when deciding between ocular extraction and eye-preserving treatment.³¹ Our study found that entropy and kurtosis were significantly correlated with TP, whereas kurtosis was positively correlated with LBD, suggesting that ADC parameters can contribute to prognostic assessment and treatment selection. These findings emphasize that enhanced intratumor heterogeneity often indicates a bigger tumor size and worse prognosis, as reported by Zhang et al.³² 

Previous studies have shown that Ki-67 and monosomy 3 are associated with poor prognosis in UM.^33–36 There is sufficient evidence that ADC histogram parameters can reflect the proliferative activity of tumor tissues.^37,38 In our study, ADC_mean, ADC_50%, and skewness were significantly correlated with the Ki-67 index. In line with our findings, prior research on patients with glioma has also demonstrated increased skewness of the ADC histogram, indicating continued tumor proliferation.³⁹ One previous study found that the mean and median ADC values are related to the composition and degree of crowding of the extracellular matrix.⁴⁰ With respect to our finding of a negative correlation between the ADC_mean and ADC_50% and the rate of tumor cell proliferation, it may be due to the fact that cell proliferation leads to an increase in cell density, which, in turn, causes the tumor tissues to be more compact and the diffusion of water molecules to be more restricted.⁴¹ In contrast, entropy was positively correlated with the percentage of monosomy 3 in the current study. This observation suggests a high degree of internal tumor heterogeneity consistent with an elevated percentage of monosomy 3. These results are encouraging for clinical practices associated with UM. Ki-67 is an important pathological parameter characterizing the degree of mitotic activity in malignant tumors, and monosomy 3 is strongly correlated with UM metastasis.⁴² However, obtaining any of these indicators requires invasive pathologic examination. Our study points to a potential noninvasive method utilizing the possible percentage of Ki-67 and monosomy 3 as ADC histogram parameters, thus providing guidance for UM risk stratification, treatment decisions, and follow-up. A smaller ADC_50% and ADC_mean and a larger skewness tend to suggest that the tumor has higher proliferative activity and is likely to grow faster or invade surrounding tissues. When entropy is larger, it suggests that patients may have a higher proportion of monosomy 3 and be more prone to metastasis. Due to the current sample size limitation, the correlations were not very strong, but these results can still serve as an important reference for non-surgical patients or before pathologic results are available. 

Patients with preoperative RD experience have an overall poorer visual prognosis. However, we found no correlation between ADC histogram parameters and preoperative RD. This finding suggests that the ADC histogram has limited value in predicting visual prognosis. 

Our study has a few limitations. It is limited by its retrospective nature, small sample size, and radiologic analyses. To control the quality of MRI imaging data, we included only orbital MRI examinations performed at our center on patients with UM. Thus, only a limited number of patients could be enrolled. Due to the small proportion of included patients, the ROC curves were not adequately smooth. Also, it was challenging to perform multifactorial analysis to determine the combined clinical significance of all of the parameters. In the future, we will expand the sample size through a multicenter approach. With a sufficient sample size, performing multifactorial analyses will increase the depth of analysis and potentially illuminate the interplay between ADC histogram parameters and clinicopathologic features. In addition, a prospective cohort will be used to enhance statistical power and to better predict prognoses. An in-depth molecular analysis, including the association between ADC parameters and molecular alterations such as BAP1 loss and GNAQ and GNA11 mutations, could further enhance the clinical significance, offering a pathway toward personalized prognostic strategies in UM. In addition, by combining MRI and other examination tools, we may be able to provide more valuable information for the diagnosis and treatment of UM. For radiologic analysis, DWI was performed on a 3T MRI machine with parameters of slice thickness of 3 mm and a slice gap of 0.5 mm. Due to the relatively small size of the tumors, only one or two slices were available for some of the tumors. Considering the current scanning parameters and tumor size limitations, it is difficult to completely avoid the effect of partial voluming, which may cause smaller tumors to be more susceptible to the ADC values of the surrounding tissues. For this reason, we excluded the voxels at the tumor border when drawing the ROIs. Also, we did not include the commonly used ADC maximum and minimum parameters in our statistical analyses, as they are extremely susceptible to the influence of surrounding tissues. Finally, the EPI readout is sensitive to B₀ inhomogeneities. The use of the EPI technique to read the signal accelerates the scanning speed, but it tends to accumulate phase differences, making the DWI image prone to magnetic susceptibility artifacts at magnetic field inhomogeneities. 

Conclusions

In conclusion, our study, to the best of our knowledge, is the first to apply ADC histograms in the preoperative imaging assessment of UM. Our results confirm the potential of ADC histogram parameters to serve as important prognostic imaging markers. Moreover, the ADC histogram parameters were found to be significantly correlated with survival prognostic factors, including tumor size, Ki-67 index, and monosomy 3, whereas they offered only limited predictive value for visual prognosis. Thus, this study presents a new noninvasive imaging tool to assist clinicians in identifying high-risk UM subtypes with poor survival prognosis at an early stage and devising optimal treatment strategies. 

Acknowledgments

The authors thank Hui Wang for her guidance in the statistical analyses. The authors thank Bullet Edits Limited for the linguistic editing and proofreading of the manuscript. 

Supported by grants from the Shanghai Science and Technology Commission (20DZ2270800, 23ZR1438400); Shanghai Eye Disease Research Center (2022ZZ01003); Rare Disease Registration Platform of Shanghai Ninth People's Hospital (JYHJB202305); Cross-Disciplinary Research Fund of Shanghai Ninth People's Hospital (JYJC202210, YG2023QNB15); and the Innovative Research Team of High-Level Local Universities in Shanghai (SHSMUZDCX20210900, SHSMU-ZDCX20210902). 

Disclosure: Y. Zheng, None; Y. Tang, None; Y. Yao, None; T. Ge, None; H. Pan, None; J. Cui, None; Y. Rao, None; X. Tao, None; R. Jia, None; S. Ai, None; X. Song, None; A. Zhuang, None