Abstract

Objectives

This study aimed to investigate the ability of radiomics features extracted from magnetic resonance imaging (MRI) images to differentiate between responders and non-responders for hepatocellular carcinoma (HCC) cases who received Y-90 transarterial radioembolization treatment.

Methods

Thirty-six cases of HCC who underwent MRI scans after Y-90 radioembolization were included in this study. Tumors were segmented from MRI T2 images, and then 87 radiomic features were extracted through the LIFEx package software. Treatment response was determined 9 months after treatment through the modified response evaluation criteria in solid tumours (mRECIST).

Results

According to mRECIST, 28 cases were responders and 8 cases were non-responders. Two radiomics features, “Grey Level Size Zone Matrix (GLSZM)-Small Zone Emphasis” and “GLSZM-Normalized Zone Size Non-Uniformity”, were the radiomics features that could predict treatment response with the area under curve (AUC)= 0.71, sensitivity= 0.93, and specificity= 0.62 for both features. Whereas the other 4 features (kurtosis, intensity histogram root mean square, neighbourhood gray-tone difference matrix strength, and GLSZM normalized grey level non-uniformity) have a relatively lower but acceptable discrimination ability range from AUC= 0.6 to 0.66.

Conclusion

MRI radiomics analysis could be used to assess the treatment response for HCC cases treated with Y-90 radioembolization.

Introduction

Hepatocellular carcinoma (HCC) stands as a significant contributor to cancer-related mortality worldwide, particularly prevalent in Asia and Africa (1). Yttrium-90 (Y-90) radioembolization emerges as a potent therapeutic modality for unresectable HCC, enabling precise delivery of high-dose of beta-radiation to the tumor while sparing adjacent healthy tissue (2). However, response to Y-90 radioembolization exhibits variability across patients, underscoring the importance of accurate response assessment in gauging treatment efficacy and guiding subsequent clinical decisions (3). Current response assessment criteria, such as modified response evaluation criteria in solid tumors (mRECIST), rely primarily on changes in tumor size, often falling short of capturing the nuanced treatment (4).

Radiomics, an advancing domain within medical imaging, presents a promising avenue for extracting quantitative features from medical images, potentially enhancing treatment response assessment. Moreover, radiomics delves into subtle nuances in image texture, shape, and intensity, imperceptible to the naked eye, yet likely correlated with tumor biology and treatment response (5). Notably, magnetic resonance imaging (MRI) radiomics has featured prominently in various studies predicting treatment responses in HCC patients. While some researchers have explored the utility of MRI-derived radiomics features in prognosticating response to transarterial chemoembolization (TACE) (6-10), few have examined their role in assessing response to transarterial radioembolization employing Y-90 in HCC patients (11, 12). The aim of this current study is to investigate the efficacy of radiomic features extracted from MRI images in discriminating between responders and non-responder to Y-90 radioembolization in HCC patients.

Materials and Methods

Patients

A cohort of 36 patients with HCC (6 females and 30 males) were included in this study. Patients aged from (37 to 87) years with average age of 67.43 years. The patients’ characteristics were shown in Figure 1. This study protocol was reviewed and approved by the Ethics Committee of Taipei Veterans General Hospital (IRB no: 2020-04-013BC, date: 13.04.2020). Informed consent was obtained from all subjects during inclusion in the study. Informed consent was obtained from all individual participants included in the study.

Imaging techniques

Planning Angiography and Technetium- 99m Macroaggregated Albumin (Tc-MAA) Single Photon Emission Computed Tomography (SPECT)/(CT)

All patients underwent a diagnostic angiogram and administration of 185 MBq (5 mCi) of Tc-MAA, followed by planar and SPECT/CT acquisitions on a hybrid SPECT/CT scanner (GE Discovery NM/CT670, USA) within one hour after Tc-MAA injection. Tc-MAA images were reconstructed on a GE Xeleris 3 workstation. The prescribed activity of the Y-90 microsphere was determined by medical internal radiation dosimetry (MIRD) model to reach a 120 Gy average dose in the target region for the glass microspheres and by partition model to reach tumor dose of 120 Gy.

Y-90 Radioembolization and Y-90 PET/MRI

All Y-90 treatments were done within 10 days after planning angiography using Y-90 glass microspheres (TheraSphere^®; Boston Scientific Corp.) or Y-90 resin microspheres (SIR-Spheres^®; Sirtex Medical Ltd.). After treatment, the post-Y90 internal pair production PET was obtained on a GE hybrid SIGNA PET/MRI with a maximum of two bed positions and 20-min acquisition per bed. MRI-T2 sequence was performed with 2D method, slice sequence= 6 mm, repetition time= 10000 msec, echo time= 106.88 msec, fat saturation= 3. All these scans and treatments occurred at our hospital between March 2018 and December 2021.

Radiomics Features

Images loaded to LIFEx Package versions 7.3.0 www.lifexsoft.org (13). Then images were segmented through three-dimensional semiautomatic tools in the package, and then 87 radiomic features were extracted, including 31 histogram features, 24 grey-level co-occurrence matrix Grey Level Cooccurrence Matrix (GLCM) features, 11 neighbourhood grey-level different matrix (GLRLM) features, 5 features grey-level run length matrix neighbourhood gray-tone difference matrix (NGTDM) features, and 16 Grey Level Size Zone Matrix (GLSZM) features. Figure 2 shows examples of HCC cases before and after segmentation.

Objective Response Rate (ORR)

The localized tumour response was defined as the response or progression within the radioembolization-treated liver. The response was evaluated using the mRECIST for HCC (14). Based on the best response of tumours observed on contrast-enhanced MRI or CT within 9 months after radioembolization. The ORR was defined as the sum of the complete response and the partial response. Eight cases were non-responders, while 28 cases were responders to treatment.

Statistical Analysis

The study used the Mann-Whitney U test to compare the responder and non-responder groups. The area under the curve from the receiver operating characteristic (ROC) was used to determine which features are sensitive to differentiating between responders and non-responders. The significance value for the tests was set at 0.05. All statistical tests were performed using Government of National Unity Public Social Private Partnership version 1.6.0.

Results

According to the Mann-Whitney U test, the only features with a significant difference between responders and non-responders had a significance value of 0.02 (Table 1), whereas most of the features had a significance value higher than 0.05, indicating that there was no significant difference between the two groups, as shown in Table 2. The area under curve (AUC) was significant for two features, “GLSZM-Small Zone Emphasis” and “GLSZM-Normalized Zone Size Non-Uniformity”, with an AUC of 0.71, and sensitivity= 0.93 and specificity= 0.62 for both features, as presented in Table 3 and Figure 3. Four features, kurtosis, intensity histogram root mean square, NGTDM strength, and GLSZM normalized grey level non-uniformity, had relatively lower but acceptable discrimination ability, with AUCs of 0.64, 0.66, 0.6, and 0.61, respectively.

The optimal cut-off values for the two highest features to distinguish between responders and non-responders were obtained using the maximum Youden index (sensitivity + specificity- 1), as shown in Table 3. Figure 4 shows a boxplot of the significant values for the optimal cut-off values obtained using the maximum Youden index.

Discussion

Several studies have explored the radiomic features of HCC patients undergoing Y-90 radioembolization, akin to our investigation. A study by Aujay et al. (11) examined 22 cases of HCC post-Y-90 radioembolization, identifying 14 patients as non-responders and 8 as responders. They extracted 107 radiomic features from arterial-phase and portal-venous phase MRI images. Their ROC analysis highlighted four radiomic parameters (long run emphasis, minor axis length, surface area, and grey level non-uniformity on arterial phase images) as predictors of early response. However, none of the statistically significant features from their study aligned with our findings. Conversely, İnce et al. (12) conducted a study with a larger cohort comprising 82 HCC patients (65 responders and 17 non-responders). They analyzed 1128 radiomic features extracted from pretreatment contrast-enhanced T1-weighted MRI scans obtained within three months before Y-90 treatment. Employing machine learning models, they identified eight radiomic features (including four first-order features, such as kurtosis, three GLCM features, and one shape feature) as optimal predictors for treatment response. Interestingly, one significant feature from their study (Kurtosis) corroborates our findings. Table 4 presents a comparison between previous studies and our current investigation.

Other studies have explored the Y-90 treatment response for HCC or other liver malignancies using diverse imaging modalities(15-17). Reimer et al. (15) assessed the efficacy of texture analysis based on post-treatment MRI of liver metastases in 37 patients to predict response to Y-90 radioembolization during follow-up for colon cancer. They exclusively utilized first-order histogram features, with Kurtosis being the sole feature aligning with our current study. Blanc-Durand et al. (16) employed whole-liver radiomics to devise a scoring system predicting progression-free survival and overall survival in unresectable HCC patients undergoing Y-90 radioembolization. Utilizing 39 imaging features, they developed a two-predictive scoring system categorizing HCC patients into low- and high-risk subgroups in a retrospective cohort of 47 patients. Key radiomics features in their predictive model included variance and NGTDM strength, mirroring aspects of our study. In a separate study, Wei et al. (17) utilized a cohort of 30 patients with primary and secondary liver tumors treated with Y-90 and employed radiomics derived from Y-90 PET to predict treatment response. They identified 15 significant features, including NGTDM strength, consistent with our findings.

Discrepancies in features may stem from variations in sample size and the statistical analysis methodology. Additionally, radiomics holds promise in enhancing medical diagnosis, yet studies have uncovered variability in radiomics values due to several factors. These factors include the voxel size (18, 19), the algorithms used for reconstruction (20, 21), the methods used for tumor segmentation (22, 23), and the discretization of grey-level values (24, 25). This variability in radiomics values can pose a challenge to the accuracy and reliability of radiomics-based diagnosis.

Study Limitations

The current study has certain limitations worth noting, such as a relatively small sample size, a restricted number of radiomic features analyzed, and absence of a comparison between our findings and other clinical information, such as histopathology results. Furthermore, we did not correlate the radiomics results with the clinical progression of the disease or liver function parameters. However, it’s essential to emphasize that the primary objective of this study was to explore the predictive capacity of radiomics in determining Y-90 treatment response, with plans to expand upon these findings in our future research endeavors.

Conclusion

In conclusion, this study shows that radiomics extracted from MRI-T2 images could be used as a non-invasive tool for predicting the response of HCC to Y-90 treatment. Two radiomic features, “GLSZM-Small Zone Emphasis” and “GLSZM-Normalized Zone Size Non-Uniformity”, were found to have the highest discrimination ability for differentiating responders from non-responders.