Highlights
-
•
A total of 4 intratumoral habitats were segmented based on CT images using otsu clustering for HCC.
-
•
Habitat radiomics outperformed traditional radiomics in stratifying prognosis for HCC.
-
•
Distinct immune status in TME contributed to the prognostic power of the habitat radiomic model.
Keywords: Hepatocellular carcinoma, Computer tomography, Habitat radiomics, Prognosis, Immune status
Abstract
Background and objective
Though several clinicopathological features are identified as prognostic indicators, potentially prognostic radiomic models are expected to preoperatively and noninvasively predict survival for HCC. Traditional radiomic models are lacking in a consideration for intratumoral regional heterogeneity. The study aimed to establish and validate the predictive power of multiple habitat radiomic models in predicting prognosis of hepatocellular carcinoma (HCC).
Methods
A total of 232 HCC patients were retrospectively included, including a training/validation cohort and two external testing cohorts from 4 centers. For habitat radiomics, intratumoral habitat partitioning based on CT images was first performed by using Otsu thresholding method. Second, a total of 350 habitat radiomic models were constructed to select the optimal model. Then, both ROC curve analyses and Kaplan-Meier survival curve analyses were applied to assess the predictive performances. Ultimately, an immune status profiling was conducted based on bioinformatic analyses and multiplex immunohistochemistry (mIHC) assays to reveal the potential mechanisms.
Results
A total of 4 habitats were segmented, and the corresponding habitat radiomic models were constructed based on each habitat and an integration of all the four habitats. Generally, habitat radiomic models outperformed traditional radiomic models in stratifying prognosis for HCC. The habitat radiomic model based on the segmented habitat 4 involving decision tree (DT) screening and random forest (RF) classifier was identified as the optimal model with an AUCmean of 0.806. Distinct resting natural killer (NK) cell infiltrations significantly contributed to the prognosis stratification of HCC by the optimal habitat radiomic model.
Conclusions
The habitat radiomic model based on CT images was potentially predictive of overall survival for HCC, with a superiority over the traditional radiomic model. The prognostic power of the habitat radiomic model was partly attributed to the distinct immune status captured in the CT images.
Introduction
As the leading primary liver malignancy, HCC is one of the main causes of cancer-related death [1]. Despite great progress in diagnosis and treatment, the overall clinical outcomes for HCC remain poor [2,3]. In view of the high heterogeneity in HCC, precise individualized treatment is expected to improve clinical management and subsequent outcomes, which is still a huge challenge right now [4]. Nevertheless, risk stratification prior to treatment is supposed to optimize the clinical benefit for HCC with personalized treatment schemes [5]. Several clinicopathological features are already identified as prognostic indicators for HCC, including alpha-fetoProtein (AFP), hepatitis virus infection, TNM staging, microvascular invasion (MVI) [6] and macrotrabecular-massive (MTM) subtype [7]. Given the statuses of MVI and MTM mainly rely on histopathologic examinations with postoperative surgical specimens, preoperatively stratifying HCC patients based on these indexes results in a dilemma. Moreover, prognostication based on AFP and TNM staging is prone to suffer from subjective evaluations. Identification of potentially prognostic imaging features is a promising approach to preoperatively and noninvasively predict survival for HCC.
Radiologic imaging, especially CT, is crucial for the routine diagnosis and treatment for HCC. Several semantic features and semi-quantitative imaging features based on plain CT and/or contrast enhanced CT (CECT) images contribute to the clinical differential diagnosis of HCC. Whereas, a comprehensive and quantitative analysis based on CT images is needed to characterize the heterogeneous nature of HCC, thus allow for non-invasive survival prognostication prior to treatment [8]. Radiomics is an emerging noninvasive analytical technique that enables to quantitatively analyze medical images in a high throughput manner [9]. Currently, increasing radiomic signature or models based on CT [10], magnetic resonance imaging (MRI) [11] and positron emission tomography/computer tomography (PET/CT) images [12] are being established to predict classification and survival outcomes for HCC. However, among all the developed radiomic models, a majority of radiomic models were developed to indirectly predict prognosis based on classification of HCC patients in terms of MVI and/or MTM [10,[12], [13], [14]]. In this study, we aimed to establish and select an optimal radiomic model based on preoperative baseline CT images to directly predict overall survival (OS) for HCC.
Though radiomics is characterized of systemically and comprehensively analysing a high throughput of quantitative features reflecting the heterogeneity in medical images, an inherent limitation of traditional radiomics is a lack of consideration for intratumoral regional heterogeneity [15]. In other words, conventional radiomics is usually carried out based on an implict assumption that the tumor is homogeneous or heterogeneous but well mixed across the whole tumor. A more sophisticated and effective strategy by incorporating intratumoral regional variation and heterogeneity quantification is expected to improve the accuracy of the predictive performance of radiomic models. Thus far, habitat radiomics which focuses on subregional radiomics with intratuoral partitioning across the whole tumor is increasingly conducted in various types of tumors, including HCC [16], breast cancer (BC) [17], non-small cell lung cancer (NSCLC) [18], nasopharyngeal carcinoma (NPC) [19] and esophageal squamous cell carcinoma (ESCC) [20].
In the present investigation, habitat radiomics based on CT images was performed to optimize the prognostic power of radiomic models for HCC. In view of the crucial role of immune status in the tumor microenvironment (TME) for prognosis of HCC [21], both bioinformatics analyses based on transcriptome sequencing data and multiplex immunohistochemistry (mIHC) assays based on HCC paraffin sections were performed to verify the association between radiomic signature and immune status, thus providing a biological explanation for the prognostic capability of these established habitat radiomic models, especially from a perspective of tumor immunity.
Materials and methods
Patient inclusion from multi-centers
Pathologically proven HCC patients (n = 232) with baseline CT imaging before surgical resection were retrospectively recruited in the study. However, HCC patients with previous history of other types of cancer or any prior treatment before CT imaging were not included in the present investigation. Additionally, HCC patients without final prognostic information or high quality CT images were also excluded. The detailed inclusion and exclusion criteria are flowcharted in Fig. 1. It is noteworthy that this is a multi-center study involving cases from four centers. Among the four centers, HCC cases (n = 101) from a public database were used as a training/validation cohort, which is named HCC-TACE-Seg (https://www.cancerimagingarchive.net/collection/hcc-tace-seg/). HCC patients (n = 82) from our institution were chosen as an external testing cohort (test 1), whereas a total of 49 HCC cases from a combination of another hospital in our city (n = 22) with TCGA-LIHC (n = 27) (https://www.cancerimagingarchive.net/collection/tcga-lihc/) were used as another external testing cohort (test 2). The clinicopathological characteristics of these included HCC patients are summarized in Table 1. Based on the follow-up information after operation, all the included HCC patients were categorized into a high survival subgroup and a low survival subgroup according to the median survival time (169 days). The respective distribution of prognosis grouping (high vs low) in each cohort is also shown in Table 1. This retrospective study was approved by institutional ethics review committee and conducted in accordance with the declaration of Helsinki and related ethical guidelines. The written informed consent requirement was waived for all the included HCC patients.
Fig. 1.
The flowchart of inclusion and exclusion process in the retrospective study from four centers.
Table 1.
The clinicopathological characteristics of HCC cohorts included in the study.
| Parameters | Training/ Validation(n = 101) | Test 1(n = 82) | Test 2(n = 49) | P value |
|---|---|---|---|---|
| Age | 67 ± 11 | 59 ± 9 | 63 ± 9 | 0.259 |
| Sex | 0.093 | |||
| Male | 65(64.4 %) | 40(48.8 %) | 26(53.1 %) | |
| Female | 36(35.6 %) | 42(51.2 %) | 23(46.9 %) | |
| HBV infection | – | |||
| Never | NA | 19(23.2 %) | NA | |
| Current or former | NA | 63(76.8 %) | NA | |
| Liver cirrhosis | – | |||
| Never | NA | 24(29.3 %) | NA | |
| Current | NA | 58(70.7 %) | NA | |
| AFP level | 0.252 | |||
| ≥400 ng/mL | 29(28.7 %) | 31(37.8 %) | NA | |
| <400 ng/mL | 72(71.3 %) | 51(62.2 %) | NA | |
| ALT | – | |||
| ≥40 U/L | NA | 33(40.2 %) | NA | |
| <40 U/L | NA | 49(59.8 %) | NA | |
| AST | – | |||
| ≥40 U/L | NA | 27(32.9 %) | NA | |
| <40 U/L | NA | 55(67.1 %) | NA | |
| TB | – | |||
| ≥17.1 μmol/L | NA | 39(47.6 %) | NA | |
| <17.1 μmol/L | NA | 43(52.4 %) | NA | |
| BCLC stage | 0.991 | |||
| A | 11(10.9 %) | 8(9.8 %) | NA | |
| B | 23(22.8 %) | 19(23.2 %) | NA | |
| C | 65(64.4 %) | 53(64.6 %) | NA | |
| D | 2(1.9 %) | 2(2.4 %) | NA | |
| Maximum tumor diameter (cm) | 7.3 (4.2–11.2) | 5.4 (2.6–8.2) | 6.6 (3.8–10.9) | 0.128 |
| OS | <0.001 | |||
| High | 23(22.8 %) | 66(80.5 %) | 27(55.1 %) | |
| Low | 78(77.2 %) | 16(19.5 %) | 22(44.9 %) |
AFP, alpha-fetoprotein; ALT, alanine aminotransferase; AST, aspartate aminotransferase; TB, total bilirubin; BCLC, Barcelona Clinic Liver Cancer; OS, overall survival.
Generation of entropy-derived CT images based on original CT images
To generate new entropy-derived CT images based on original CT images, the local entropy of each voxel was first calculated from CT images as a measure of local texture pattern and the degree of chaos in a small neighborhood of each voxel (Fig. 2A). Higher entropy values indicate greater image complexity or randomness, while lower values suggest more uniform or structured images. Images preprocessing and volume of interest (VOI) segmentation were first performed to ensure the original CT images were suitable for analysis, which involved resampling and masking of VOI by an extension of 15 voxels at each three-dimensional direction. Then, a moving window with a patch size of 9 × 9 × 9 was used to calculate the local entropy of each voxel from the original CT images. The CT local entropy was calculated by using Python's scipy.stats package, and new CT images were finally generated based on the calculated entropy intensity of each voxel. Ultimately, for each voxel, CT value and local entropy were obtained to build a two-dimensional feature vector. The joint use of these two features enables intratumoral subregion partitioning for habitat radiomics.
Fig. 2.
The flowchart of habitat radiomics in this study. Briefly, this investigation consists of four steps. (A) To generate new entropy-derived CT images based on original CT images, the local entropy of each voxel was first calculated from CT images. Then, CT value and CT entropy were obtained to build a two-dimensional feature vector, thus enabling intratumoral subregion partitioning for habitat radiomics by using Otsu thresholding method. (B) A total of 350 (5 × 7 × 10) habitat radiomic models were established based on each habitat (n = 4) and an integration of all the four habitats involving different combination sets of 7 feature screening methods with 10 machine learning classifiers. (C) Evaluating the performances of constructed habitat radiomic models in prognosis prediction for HCC. Among the 350 constructed models, ROC curves, calibration curves and decision curves were drawn to select the optimal habitat radiomic model. Furthermore, Kaplan-Meier survival curve analyses were also conducted to verify the prognostic value of the optimal habitat radiomic model. (D) Bioinformatics analyses and multiplex immunohistochemistry (mIHC) assays to reveal the potential explanations for the prognostic habitat radiomic model.
Habitat radiomics
As known, traditional radiomics was conducted following a canonial flowchart, which consists of resampling, VOI delineation and masking (The 3D Slicer version 5.2.2; open-source software; https://www.slicer.org/), radiomics features extraction, radiomics features screening and selection, radiomics models constuction and predictive performance evaluation, utilizing open-source Python package Pyradiomics 1.2.0 (http://www.radiomics.io/pyradiomics.html). The whole VOI for each included HCC patient was manually delineated slice-by-slice based on CT images by two nuclear medicine physicians with over 5 years of experience by using the 3D Slicer software. In addition, for each physician, the VOI delineation was performed twice. Different from traditional radiomics, habitat radiomics was performed to determine the intratumoral subregion variation in heterogeneity captured by the medical images. Based on original CT images and generated entropy-derived CT images, the intratumoral habitat segmentation was first achieved by using Otsu thresholding method. Briefly, for each voxel, CT value and local entropy were obtained to build a two-dimensional feature vector. Then, a total of four habitats were segmented in each VOI by using Otsu thresholding method, which consist of four types of representive sub-regions with high CT value and high CT entropy, high CT value and low CT entropy, low CT value and high CT entropy and low CT value and low CT entropy, respectively (Fig. 2A). After habitat segmentation, habitat radiomics were then performed. In the present investigation, habitat radiomics were based on a total of 5 types of habitats (habitat 1–4 and an integration of all the four habitats). By involving combinations of 7 feature screening methods (analysis of variance (ANOVA), mutual information, Ridge, support vector machine (SVM), RF, DT and Xgboost) with 10 machine learning classifiers (SVM, DT, RF, extra tree (ET), k-nearest neighbors (KNN), Xgboost, multinomial naive Bayes, logistic regression, Gaussian process and gradient boosting), a total of 350 (5 × 7 × 10) habitat radiomic models were established to choose the optimal habitat radiomic model (Fig. 2B). A 5-fold cross validation with a repetition of 100 times were performed in the training/validation cohort during the constructions of multiple habitat radiomic models. ROC curve analyses, calibration curve analyses and decision curve analyses were conducted to assess the prognostic performances of the habitat radiomic models in predicting survival for HCC. Kaplan-Meier survival curves were drawn to determine the differences in survival times between the model predicted high survival subgroup and low survival subgroup (Fig. 2C). The overall flowchart of habitat radiomics is depicted in Fig. 2A–C.
Bioinformatic analyses based on transcriptome sequencing data
Transcriptome sequencing data of the included HCC patients from an external cohort (TCGA-LIHC) were used to conduct bioinformatic analyses to identify the underlying biological mechanisms for the prognostic habitat radiomic model (Fig. 2D). First, the differentially expressed genes (DEGs) between the model predicted high survival subgroup and low survival subgroup were determined. The volcano plots were depicted by using the ggplot2 package in R software version 3.6.0 (https://www.r-project.org/). Second, both Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were performed to identify significantly enriched GO terms and KEGG pathways based on the specific gene lists by using clusterProfiler. Bubble diagrams were also drawn by using the ggplot2 package in R software. Then, a potential immunological explanation for the prognostic habitat radiomic model was detected by using a gene set enrichment analysis (GSEA) specializing in immunologic signature gene (GSEA | MSigDB (gsea-msigdb.org)). Furthermore, the composition of the infiltrating immune cells in HCC was analyzed by using CIBERSORT (https://cibersort.stanford.edu/), and violin plots were drawn to display the differences in immune cell infiltrations between the two predicted survival subgroups.
Multiplex immunohistochemistry
Paraffin-embedded sections of the included HCC cohort from our institution were used to perform mIHC assay. After antigen retrieval and blocking, the tissue sections were incubated with primary antibody (anti-CD3, 1:400, Cell Signaling Technology, Inc., MA, USA) at 4°C overnight. Then, horseradish peroxidase (HRP)-conjugated secondary antibody incubation was conducted following washing. Following an incubation at room temperature for 60 min, the slides were washed again to remove unbound secondary antibodies. For detection, an appropriate fluorescence substrate (Akoya Biosciences, Inc., MA, USA) was applied. The steps from antigen retrieval to the application of fluorescence substrate were repeated again to detect CD56 (1:200, Cell Signaling Technology, Inc., MA, USA). In the end, counterstaining tissue sections with a nuclear stain (DAPI, Thermo Fisher Scientific Inc., Shanghai, China) was carried out to visualize the cell nuclei, and the tissue sections with an appropriate mounting medium to preserve fluorescence and prevent photobleaching. To quantify the stained immune cells, a fluorescence microscope was employed to visualize and capture the signals from different fluorophores by using the inForm software version 2.6 (Akoya Biosciences, Inc., MA, USA). Infiltrating immune cells with positive staining for CD56 but negative staining for CD3 were considered as natural killer cells (NKs).
Statistical analyses
The inter-group statistical differences for numerical variables were determined by a t-test (normal distribution), whereas a chi-square test was used for categorical variables. For radiomic features screening, the intra-observer and inter-observer consistency in the habitat delineations were first assessed. Briefly, the intraclass and the interclass correlation coefficients (ICCs) were calculated, and only radiomic features with an ICC value ≥ 0.95 were deemed robust and retained, which represent stable radiomic features independent of the process of delineation. Then, a Spearman correlation analysis with a Spearman correlation coefficient of 0.75 was performed to reduce the redundancy between selected radiomic features. In the end, a total of 7 screening methods were used to identify the informative radiomic features that significantly differ between the two survival-related subgroups. In the ROC curves analyses, the Delong test was employed to identify the differences in the performances of all the developed radiomic models in prognosis prediction for HCC. In the Kaplan-Meier survival curves analyses, a log-rank test was used to detect the differences in survival outcomes between the model predicted high survival subgroup and low survival subgroup. For bioinformatic analyses, R software version 3.6.0 and related packages were used.
Results
Patient characteristics
A total of 232 HCC patients from 4 centers were retrospectively included in the study, and all the patients were categorized into a low and a high survival subgroup according to a median survival time of 169 days. Clinicopathologic characteristics and survival grouping results are listed in Table 1 in terms of one training/validation cohort and two external testing cohorts (Test 1 and Test 2). The cohort with the most reruited HCC cases (n = 101) from a public database (HCC-TACE-Seg) was used as the training/validation cohort. HCC patients in our institution (n = 82) were used as one independent external testing cohort (Test 1), and a combination of included HCC patients from TCGA-LIHC public database (n = 27) with another hospital in our city (n = 22) were used as another external testing cohort (Test 2). As shown, no statistically significant differences were found in all the clinicopathologic characteristics between different cohorts. Noticeably, the survival groupings were markedly different across different cohorts, partly due to an inevitable bias of sampling in a retrospective investigation involving multiple cohorts.
Habitat radiomics based on original CT images and generated entropy-derived CT images
The prerequisite for habitat radiomics is intratumoral subregion partitioning in the VOI of the lesion based on multimodal-like medical images. Given that only baseline plain CT images were obtained in the study, a corresponding entropy-derived CT image was first generated for each original CT image (Fig. 3A). Based on the bimodal-like CT images, Otsu thresholding method was employed to achieve intratumoral subregion partitioning in the delineated VOI. As demonstrated in Fig. 3B, four habitats are segmented in each VOI. Then, a total of 1326 radiomic features were extracted based on each habitat. After standardization, Pearson correlation analysis and ICC analyses, a set of 376, 315, 375 and 403 radiomic features were selected for habitat 1, 2, 3 and 4, respectively. Subsequently, a set of 70 (7 × 10) radiomic models involving combinations of 7 feature screening methods with 10 machine learning classifiers were established based on each habitat delineated above (Fig. 2B).
Fig. 3.
Intratumoral habitat segmentation based on original CT images and generated entropy-derived CT images. (A) Representative original CT images and generated entropy-derived CT images from two HCC cases. (B) Representative images for VOI segmentation and intratumoral habitat partitioning from tow HCC cases. A total of 4 habitats were segmented based on each VOI by using Otsu thresholding method.
Prognostic value of habitat radiomic model in survival prediction for HCC
After respectively selecting the optimal combination of the feature screening method with a machine learning classifier for each type of habitat, 5 habitat radiomic models were finally selected for further analyses, including habitat 1, 2, 3 4 and an integration of all the habitats. Detailed descriptions of the finally selected radiomic features in the aforementioned 5 habitat radiomic models are listed in Supplementary Table 1–5. As shown in Table 2, sensitivity, specificity, accuracy and AUC are used as the main indicators to assess the performances of the selected habitat radiomic models in survival prediction for HCC. Generally, each habitat radiomic model outperformed the traditional radiomic model, suggesting a superiority of habitat radiomics over conventional radiomics in prognostic power for HCC. Finally, the habitat radiomic model based on segmented habitat 4 involving DT screening and RF classifier was identified as the optimal model with an AUCmean of 0.806 in two external testing cohorts. Additionally, the ROC curves, the calibration curves and the decision curves (Fig. 4A–C) were also depicted for the optimal habitat radiomic model in one internal validation cohort and two external testing cohorts, repectively. To directly verify the prognostic power of the selected optimal habitat radiomic model in survival classification for HCC, Kaplan-Meier survival curves (Fig. 4D, E) were drawn to determine the differences in survival times between the model predicted high survival subgroup and low survival subgroup. The log-rank tests confirmed that the predicted high survival subgroup exhibited a remarkably favorable survival than the predicted low survival subgroup in both of the two external testing cohorts (P = 0.003 and P = 0.03).
Table 2.
The prognostic power of the constructed habitat radiomics models in survival prediction for HCC in comparison with that of traditional radiomics.
| Models | Feature screening | Classifiers | Sensitivity | Specificity | Accuracy | AUC | AUCmean |
|---|---|---|---|---|---|---|---|
| Traditional radiomics | DT | Xgboost | 0.673 | ||||
| Validation | 0.634 | 0.581 | 0.605 | 0.657 | |||
| Test 1 | 0.604 | 0.673 | 0.635 | 0.737 | |||
| Test 2 | 0.667 | 0.5 | 0.634 | 0.625 | |||
| Habitat radiomics 1 | DT | RF | 0.756 | ||||
| Validation | 0.7 | 0.692 | 0.699 | 0.743 | |||
| Test 1 | 0.667 | 0.818 | 0.735 | 0.835 | |||
| Test 2 | 0.606 | 0.563 | 0.598 | 0.691 | |||
| Habitat radiomics 2 | RF | RF | 0.72 | ||||
| Validation | 0.641 | 0.569 | 0.586 | 0.695 | |||
| Test 1 | 0.641 | 0.718 | 0.676 | 0.754 | |||
| Test 2 | 0.621 | 0.688 | 0.634 | 0.71 | |||
| Habitat radiomics 3 | SVM | KNN | 0.639 | ||||
| Validation | 0.526 | 0.53 | 0.534 | 0.642 | |||
| Test 1 | 0.541 | 0.572 | 0.555 | 0.627 | |||
| Test 2 | 0.636 | 0.5 | 0.61 | 0.647 | |||
| Habitat radiomics 4 | DT | RF | 0.806 | ||||
| Validation | 0.784 | 0.687 | 0.737 | 0.805 | |||
| Test 1 | 0.704 | 0.818 | 0.755 | 0.865 | |||
| Test 2 | 0.818 | 0.563 | 0.768 | 0.747 | |||
| Integrated habitat | DT | RF | 0.709 | ||||
| Validation | 0.745 | 0.597 | 0.66 | 0.747 | |||
| Test 1 | 0.567 | 0.711 | 0.635 | 0.756 | |||
| Test 2 | 0.697 | 0.438 | 0.646 | 0.624 |
DT, decision tree; RF, random forest; SVM, support vector machine; KNN, k-nearest neighbors.
Fig. 4.
Evaluation of the performace of the optimal habitat radiomic model based on CT images in prognosis prediction for HCC. (A) The ROC curves, the calibration curves and the decision curves were depicted for the optimal habitat 4-based radiomic model in one internal validation cohort and two external testing cohorts, respectively. (B) Kaplan-Meier survival curves indicated significant differences in survival times between the model predicted high survival subgroup and low survival subgroup from both the external tesing cohorts (p = 0.003 and p = 0.03).
Distinct molecular profiles between different prognosis subgroups predicted by the optimal habitat radiomic model
To reveal the underlying molecular biological mechanism responsible for the prognostic value of the optimal habitat radiomic model, radiotranscriptomic analysis was performed in the study (Fig. 5A). As indicated in the volcano plots (Fig. 5B), the association between the predicted prognosis grouping and molecular profile are determined, which are reflected in numerous differently expressed genes (DEGs) between the model predicted high survival subgroup and low survival subgroup. Generally, these DEGs were mainly composed of key molecules in carcinogenesis, signal transduction pathways and interaction between cells, such as CTAG2, EREG, CDH17, ZBTB38, IGSF3, GPRC5A, FOXN4, ROBO2 and SYT1. Consistently, the GO enrichment analyses (Fig. 5C) and KEGG enrichment analyses (Fig. 5D) also confirmed that multiple significant biological processes contributed to the model predicted prognosis grouping, including cellular response to xenobiotic stimulus, drug metabolism, nucleotide metabolism and neurotransmitter transport.
Fig. 5.
Distinct molecular profiles between model predicted high survival and low survival subgroups by using bioinformatic analysis based on transcriptome sequencing data. (A) Scheme for the radiotranscriptomic analysis. (B) Numerous DEGs between the predicted high survival and low survival subgroups were shown in the volcano plots. Bubble plots of the GO enrichment analysis (C) and KEGG enrichment analysis (D) between the model predicted high survival and low survival subgroups in HCC.
Determination and validation of immune status based on prognosis subgroup classification predicted by the optimal habitat radiomic model
Immune status in TME was previously reported as a prognostic indicator in HCC. To disclose a potential immunological mechanism underlying the prognostic habitat radiomic model, the CIBERSORT method was used to characterize the differences in tumor-infiltrating immune cells between the model predicted high survival subgroup and low survival subgroup. The violin plots in Fig. 6A show that the fractions of resting NKs are statistically different between the two subgroups. Additionally, a GSEA analysis specializing in immunologic signatures was also conducted. As illustrated in Fig. 6B, the level of regulation of NK cell mediated immunity is proven to be closely related to the model predicted prognosis subgroup classification. To verify the bioinformatic conclusion in HCC tissue sections, mIHC assays were carried out in testing cohort 1. As expected, the proportions of infiltrating resting NKs in the model predicted high survival subgroup was decreased in comparison with that in the model predicted low survival subgroup (Fig. 6C).
Fig. 6.
Determination and validation of immune status distinctions between model predicted high survival and low survival subgroups. (A) Bioinformatic analysis by using CIBERSORT was carried out to determine the differences in immune cell infiltations between the model predicted high survival and low survival subgroups. (B) A gene set enrichment analysis (GSEA) specializing in immunologic signature was also conducted to identify the potential immunological mechanism for the prognostic habitat radiomic model. (C) mIHC tests were performed to verify the conclusion obtained from bioinformatic analysis by using CIBERSORT. (Red, CD3+T cells; Green, CD56+ natural killer cells; Yellow, CD3+CD56+T cells).
Discussion
In the present investigation, a variety of habitat radiomic models based on CT images were established to predict the prognosis for HCC with a multi-cohort validation. Among various constructed models, a model based on habitat 4 involving DT screening and RF classifier was identified as the optimal habitat radiomic model, with superiority over traditional radiomic model in predicting survival for HCC. Habitat radiomic model was promisingly used as an effective and noninvasive tool to preoperatively stratify HCC patients based on prognosis prediction, thus optimizing the clinical benefit for HCC by administration of individualized treatment. Furthermore, distinct resting NK cell infiltrations in TME were revealed in different model predicted survival subgroups, suggesting a potential immunological mechanism in part responsible for the prognostic capability of the habitat radiomic model.
Though radiomic models to predict prognosis for HCC were previously reported [10,12,[22], [23], [24]], studies regarding habitat radiomics based on intratumoral subregion segmentation is currently few. As known, traditional radiomics were based on an assumption of globally well mixed intratumoral heterogeneity across the whole tumor, whereas habitat radiomics explicitly characterize the subregional variations in heterogeneity based on intratumoral habitat partitioning. Originally, habitat radiomics were conducted based on multimodal medical images or multimodal-like medical images, such as dynamic CECT [20,25,26], multi-sequence MRI [16,27] and PET/CT [18,19,28,29]. In the study, the generation of entropy-derived CT images based on original plain CT images enables us to conduct habitat radiomics depending on single modal imaging, which is a remarkable innovation in the study. Additionally, comprehensive and systematic habitat radiomic analyses were performed based on different combination sets of 7 feature screening methods with 10 machine learning classifiers, aiming to select the optimal habitat radiomic model, which is another noteworthy novelty of this investigation.
With the increasing development of radiomics and habitat radomics, potential biological explanations for the highly predictive or prognostic radiomic models were urgently needed to achieve their translational application in clinical practice [30]. The booming artificial intelligence (AI) techniques in multiple fields lead to an increasing number of comprehensive and systematic multi-omics studies to provide more information regarding tumor heterogeneity, including radiomics, genomics, transcriptomics, proteomics and metabonomics. In particular, radiotranscriptomics which combines radiomics and transcriptomics is increasingly employed to identify potential biological explanations for constructed predictive or prognostic radiomic models [25,31,32]. As known, the immune landscape in the TME significantly contributes to carcinogenesis, progression and survival outcomes in HCC [21,33,34], and radiomic models for the prediction of immune status were also constructed and validated in previous reports [13,31,35,36]. A bioinformatic analysis based on transcriptome sequencing data uncovered a variety of DEGs and distinct biological processes between the model predicted high survival subgroup and low survival subgroups in the study. Particularly, distinct resting NK cell infiltrations in TME between the two predicted subgroups were identified and verified by mIHC assays, suggesting a potential immunological explanation for the prognostic habitat radiomic model. Consistently, the status of NK cell infiltration was previously proven to be a vital indicator for the prognosis of HCC [[37], [38], [39]]. In the present investigation, a decreased infiltration of resting NK cells in the predicted high survival subgroup in contrast with that in the low survival subgroup was found, suggesting an enhanced level of activated innate immunity in the model predicted high survival subgroup compared to that in the model predicted low survival subgroup. An appropriate initiation and exertion of immune response are expected to efficiently confront the tumor, and thus resulting in a favorable prognosis. As aforementioned, habitat radiomics is superior to traditional radiomics in quantitative characterization of intratumoral regional heterogeneity variation. However, more advanced single cell spatial transcriptomics and single cell spatial proteomics are warranted in future studies to reveal the underlying biological mechanisms for this established prognostic habitat radiomic model, though some immunological mechanisms were identified via bioinformatic analyses and mIHC assays in this study [[40], [41], [42]].
Despite the inspiring results concerning the outperformance of habitat radiomic model over traditional radiomic model in survival prediction for HCC, several limitations of this study are needed to be addressed. First, this is a retrospective study with a limited sample size, thus a sample bias is inevitable in the present investigation. Nevertheless, the design of multi-cohort validation in part alleviated the inherent deficiency. Undoubtedly, a prospectively designed study with a larger sample size involving multi-institution is warranted to corroborate the obtained conclusion in future study. Second, only original plain CT images and generated entropy-derived CT images were used for habitat radiomics in the study. To verify the superiority of habitat radiomic models in survival prediction for HCC, habitat radiomics based on alternative imaging modalities, such as CECT, MRI and PET/CT, were expected in future study. Then, except for Otsu thresholding method, more subregion segmentation methods, such as some machine learning or deep learning based segmentation techniques should be chosen in habitat radiomics. In the end, the underlying biological explanation for the prognostic habitat radiomic models was limited to bioinformatic analyses based on transcriptome sequencing data, more comprehensive and systematic multi-omics investigations are suggested to identify and verify more potential mechanisms.
Conclusion
To sum up, the generation of entropy derived CT images based on original CT images allowed for habitat radiomics to predict prognosis of HCC, which exhibited an outperformance in comparison with traditional radiomic models. Radiotranscriptomics revealed a relationship between immune status in TME and prognosis classification predicted by CT-based habitat radiomic models, suggesting a potential immunological explanation for the prognostic power of the constructed optimal habitat radiomic model.
CRediT authorship contribution statement
Kun Chen: Writing – original draft, Methodology, Investigation, Formal analysis, Conceptualization. Chunxiao Sui: Writing – original draft, Methodology, Investigation, Formal analysis. Ziyang Wang: Methodology, Investigation, Formal analysis. Zifan Liu: Investigation. Lisha Qi: Writing – review & editing, Resources. Xiaofeng Li: Writing – review & editing, Supervision, Resources, Funding acquisition, Conceptualization.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgement
This work is supported by grants from the National Natural Science Foundation of China (82272074) and Tianjin Key Medical Discipline (Specialty) Construction Project (TJYXZDXK-009A).
Footnotes
Supplementary material associated with this article can be found, in the online version, at doi:10.1016/j.tranon.2024.102260.
Contributor Information
Lisha Qi, Email: lqi01@tmu.edu.cn.
Xiaofeng Li, Email: xli03@tmu.edu.cn.
Appendix. Supplementary materials
Data Availability
The original codes used in the study is provided as supplementary materials for verification. The raw data used and/or analyzed in the study from two institutions in Tianjin are available from the corresponding author on reasonable request. For other raw data used and/or analyzed in the study, please refer to the corresponding public database.
References
- 1.Singal A.G., et al. Global trends in hepatocellular carcinoma epidemiology: implications for screening, prevention and therapy. Nat. Rev. Clin. Oncol. 2023;20:864–884. doi: 10.1038/s41571-023-00825-3. [DOI] [PubMed] [Google Scholar]
- 2.Rimassa L., et al. Combination immunotherapy for hepatocellular carcinoma. J. Hepatol. 2023;79:506–515. doi: 10.1016/j.jhep.2023.03.003. [DOI] [PubMed] [Google Scholar]
- 3.Jiang C., et al. Radiomics in the diagnosis and treatment of hepatocellular carcinoma. Hepatobiliary Pancreat. Dis. Int. 2023;22:346–351. doi: 10.1016/j.hbpd.2023.03.010. [DOI] [PubMed] [Google Scholar]
- 4.Yang X., et al. Precision treatment in advanced hepatocellular carcinoma. Cancer Cell. 2024;42:180–197. doi: 10.1016/j.ccell.2024.01.007. [DOI] [PubMed] [Google Scholar]
- 5.Lee Y.T., et al. Risk stratification and early detection biomarkers for precision HCC screening. Hepatology. 2023;78:319–362. doi: 10.1002/hep.32779. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Yao L.Q., et al. Grading severity of microscopic vascular invasion was independently associated with recurrence and survival following hepatectomy for solitary hepatocellular carcinoma. Hepatobiliary Surg. Nutr. 2024;13:16–28. doi: 10.21037/hbsn-22-411. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Matsuura T., et al. Histological diagnosis of polyploidy discriminates an aggressive subset of hepatocellular carcinomas with poor prognosis. Br. J. Cancer. 2023;129:1251–1260. doi: 10.1038/s41416-023-02408-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Ronot M., et al. Imaging to predict prognosis in hepatocellular carcinoma: current and future perspectives. Radiology. 2023;307 doi: 10.1148/radiol.221429. [DOI] [PubMed] [Google Scholar]
- 9.Mayerhoefer M.E., et al. Introduction to Radiomics. J. Nucl. Med. 2020;61:488–495. doi: 10.2967/jnumed.118.222893. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Xu X., et al. Radiomic analysis of contrast-enhanced CT predicts microvascular invasion and outcome in hepatocellular carcinoma. J. Hepatol. 2019;70:1133–1144. doi: 10.1016/j.jhep.2019.02.023. [DOI] [PubMed] [Google Scholar]
- 11.Liu H.F., et al. Multiparametric MRI-based intratumoral and peritumoral radiomics for predicting the pathological differentiation of hepatocellular carcinoma. Insights Imaging. 2024;15:97. doi: 10.1186/s13244-024-01623-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Li Y., et al. Radiomics analysis of [(18)F]FDG PET/CT for microvascular invasion and prognosis prediction in very-early- and early-stage hepatocellular carcinoma. Eur. J. Nucl. Med. Mol. Imaging. 2021;48:2599–2614. doi: 10.1007/s00259-020-05119-9. [DOI] [PubMed] [Google Scholar]
- 13.Feng Z., et al. CT radiomics to predict macrotrabecular-massive subtype and immune status in hepatocellular carcinoma. Radiology. 2023;307 doi: 10.1148/radiol.221291. [DOI] [PubMed] [Google Scholar]
- 14.Xia T.Y., et al. Predicting microvascular invasion in hepatocellular carcinoma using CT-based radiomics model. Radiology. 2023;307 doi: 10.1148/radiol.222729. [DOI] [PubMed] [Google Scholar]
- 15.Zhang L., et al. The progress of multimodal imaging combination and subregion based radiomics research of cancers. Int. J. Biol. Sci. 2022;18:3458–3469. doi: 10.7150/ijbs.71046. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Gao L., et al. Multi-region radiomic analysis based on multi-sequence MRI can preoperatively predict microvascular invasion in hepatocellular carcinoma. Front. Oncol. 2022;12 doi: 10.3389/fonc.2022.818681. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Jiang T., et al. Intratumoral analysis of digital breast tomosynthesis for predicting the Ki-67 level in breast cancer: a multi-center radiomics study. Med. Phys. 2022;49:219–230. doi: 10.1002/mp.15392. [DOI] [PubMed] [Google Scholar]
- 18.Shen H., et al. A subregion-based positron emission tomography/computed tomography (PET/CT) radiomics model for the classification of non-small cell lung cancer histopathological subtypes. Quant. Imaging Med. Surg. 2021;11:2918–2932. doi: 10.21037/qims-20-1182. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Xu H., et al. Subregional radiomics analysis of PET/CT imaging with intratumor partitioning: application to prognosis for nasopharyngeal carcinoma. Mol. Imaging Biol. 2020;22:1414–1426. doi: 10.1007/s11307-019-01439-x. [DOI] [PubMed] [Google Scholar]
- 20.Xie C., et al. Sub-region based radiomics analysis for survival prediction in oesophageal tumours treated by definitive concurrent chemoradiotherapy. EBioMedicine. 2019;44:289–297. doi: 10.1016/j.ebiom.2019.05.023. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Donne R., et al. The liver cancer immune microenvironment: therapeutic implications for hepatocellular carcinoma. Hepatology. 2023;77:1773–1796. doi: 10.1002/hep.32740. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Nie P., et al. Incremental value of radiomics-based heterogeneity to the existing risk criteria in predicting recurrence of hepatocellular carcinoma after liver transplantation. Eur. Radiol. 2023;33:6608–6618. doi: 10.1007/s00330-023-09591-3. [DOI] [PubMed] [Google Scholar]
- 23.Fiz F., et al. PET/CT-based radiomics of mass-forming intrahepatic cholangiocarcinoma improves prediction of pathology data and survival. Eur. J. Nucl. Med. Mol. Imaging. 2022;49(10):3387–3400. doi: 10.1007/s00259-022-05765-1. [DOI] [PubMed] [Google Scholar]
- 24.Ji G.W., et al. Radiomic Features at Contrast-enhanced CT Predict Recurrence in Early Stage Hepatocellular Carcinoma: a Multi-Institutional Study. Radiology. 2020;294:568–579. doi: 10.1148/radiol.2020191470. [DOI] [PubMed] [Google Scholar]
- 25.Javed S., et al. Risk prediction of pancreatic cancer using AI analysis of pancreatic subregions in computed tomography images. Front. Oncol. 2022;12 doi: 10.3389/fonc.2022.1007990. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Xia W., et al. Radiogenomics of hepatocellular carcinoma: multiregion analysis-based identification of prognostic imaging biomarkers by integrating gene data-a preliminary study. Phys. Med. Biol. 2018;63 doi: 10.1088/1361-6560/aaa609. [DOI] [PubMed] [Google Scholar]
- 27.Zhang X., et al. Survival-relevant high-risk subregion identification for glioblastoma patients: the MRI-based multiple instance learning approach. Eur. Radiol. 2020;30:5602–5610. doi: 10.1007/s00330-020-06912-8. [DOI] [PubMed] [Google Scholar]
- 28.Pan Z., et al. A subregion-based prediction model for local-regional recurrence risk in head and neck squamous cell carcinoma. Radiother. Oncol. 2023;184 doi: 10.1016/j.radonc.2023.109684. [DOI] [PubMed] [Google Scholar]
- 29.Wu J., et al. Tumor subregion evolution-based imaging features to assess early response and predict prognosis in oropharyngeal cancer. J. Nucl. Med. 2020;61:327–336. doi: 10.2967/jnumed.119.230037. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Tomaszewski M.R., et al. The biological meaning of radiomic features. Radiology. 2021;298:505–516. doi: 10.1148/radiol.2021202553. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Aoude L.G., et al. Radiomics biomarkers correlate with CD8 expression and predict immune signatures in melanoma patients. Mol. Cancer Res. 2021;19:950–956. doi: 10.1158/1541-7786.MCR-20-1038. [DOI] [PubMed] [Google Scholar]
- 32.Tixier F., et al. Transcriptomics in cancer revealed by Positron Emission Tomography radiomics. Sci. Rep. 2020;10:5660. doi: 10.1038/s41598-020-62414-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Li R., et al. Identification and validation of a novel tumor microenvironment-related prognostic signature of patients with hepatocellular carcinoma. Front. Mol. Biosci. 2022;9 doi: 10.3389/fmolb.2022.917839. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Xiang S., et al. Identification of prognostic genes in the tumor microenvironment of hepatocellular carcinoma. Front. Immunol. 2021;12 doi: 10.3389/fimmu.2021.653836. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Dercle L., et al. Emerging and evolving concepts in cancer immunotherapy imaging. Radiology. 2023;306:32–46. doi: 10.1148/radiol.210518. [DOI] [PubMed] [Google Scholar]
- 36.Wang X., et al. Radiomics predicts the prognosis of patients with locally advanced breast cancer by reflecting the heterogeneity of tumor cells and the tumor microenvironment. Breast Cancer Res. 2022;24:20. doi: 10.1186/s13058-022-01516-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Li M., et al. The novel-natural-killer-cell-related gene signature predicts the prognosis and immune status of patients with hepatocellular carcinoma. Int. J. Mol. Sci. 2023;24(11):9587. doi: 10.3390/ijms24119587. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Li S., et al. Integrated analysis of single-cell and bulk RNA-sequencing reveals tumor heterogeneity and a signature based on NK cell marker genes for predicting prognosis in hepatocellular carcinoma. Front. Pharmacol. 2023;14 doi: 10.3389/fphar.2023.1200114. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39.Feng Q., et al. Combining bulk and single-cell RNA-sequencing data to develop an NK cell-related prognostic signature for hepatocellular carcinoma based on an integrated machine learning framework. Eur. J. Med. Res. 2023;28:306. doi: 10.1186/s40001-023-01300-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40.Liu Y., et al. Identification of a tumour immune barrier in the HCC microenvironment that determines the efficacy of immunotherapy. J Hepatol. 2023;78:770–782. doi: 10.1016/j.jhep.2023.01.011. [DOI] [PubMed] [Google Scholar]
- 41.Wu L., et al. An invasive zone in human liver cancer identified by Stereo-seq promotes hepatocyte-tumor cell crosstalk, local immunosuppression and tumor progression. Cell Res. 2023;33:585–603. doi: 10.1038/s41422-023-00831-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 42.Wu R., et al. Comprehensive analysis of spatial architecture in primary liver cancer. Sci Adv. 2021;7:eabg3750. doi: 10.1126/sciadv.abg3750. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The original codes used in the study is provided as supplementary materials for verification. The raw data used and/or analyzed in the study from two institutions in Tianjin are available from the corresponding author on reasonable request. For other raw data used and/or analyzed in the study, please refer to the corresponding public database.