DOI: https://doi.org/10.21203/rs.3.rs-2055732/v1
Using texture features derived from contrast-enhanced computed tomography (CT) combined with general imaging features as well as clinical information to predict treatment response and survival in patients with hepatocellular carcinoma (HCC) who received transarterial chemoembolization (TACE) treatment.
From January 2014 to June 2020, 301 patients with HCC who underwent TACE were retrospectively reviewed. Their treatment-naïve contrast-enhanced CTs were retrieved and reviewed by two independent radiologists. Four general imaging features were evaluated, including the largest lesion diameter, the number of lesions, the presence or absence of portal vein thrombus and the presence or absence of ascites. Texture features were extracted based on the regions of interest (ROIs) drawn on the slice with the largest axial diameter of all lesions using Pyradiomics. After excluding features with low reproducibility, the remaining features with high prognostic value were selected for further analyses. The data were randomly divided in a ratio of 8:2 for model training and testing. A random forest classifier was built to predict treatment response. Random survival forest and Cox proportional-hazards models were constructed to predict patients’ overall (OS) and progress-free survival (PFS).
We retrospectively evaluated 301 patients (average 55.3 years old) with HCC treated with TACE. After feature reduction and selection, 22 features were included in model construction. The random forest classifier including texture features achieved an area under the curve (AUC) of 0.968 with an accuracy of 88.3% for predicting treatment response, significantly higher than the model without texture features. Along with important clinical and general image features, texture features are significantly correlated with overall and progress-free survival, especially for Gray-level size zone matrix (GLSZM) group (p < 0.05).
Random forest algorithm based on texture features combined with general imaging features, and clinical information is a robust method for predicting prognosis in patients with HCC treated with TACE, which may help avoid additional examinations and assist in treatment planning.
Hepatocellular carcinoma (HCC) is a malignant disease with high mortality. Many risk factors have been well established that impact the outcomes of HCC, including age, gender, staging, ascites, tumor thrombus and liver function [1]. Curative surgery will increase the long-term survival rate. However, not all HCCs can be treated with surgical resection due to the high disease burden, insufficient residual liver volume, severe cirrhosis, disseminated metastatic lesions within the liver, presence of portal vein tumor thrombus and other cancer-related symptoms [2].
The Barcelona Clinic Liver Cancer (BCLC) staging system supported TACE as the first treatment choice in patients with unresectable HCC, such as those with large or multinodular HCC. The same recommendation is also made in the Chinese University Prognostic Index (CUPI) [3] and the Hong Kong Liver Cancer (HKLC) staging system [4]. The long-term survival was prolonged in patients with unresectable HCC when treated with TACE compared to best supportive care [5]. However, in the clinical setting, the therapeutic outcome of TACE is not always satisfying when it comes to individual cases because the biological behavior of tumor cells is highly heterogeneous.
Currently, the assessment of TACE outcomes mainly depends on imaging methods, such as computed tomography (CT) and magnetic resonance imaging (MRI). Modified Response Evaluation Criteria In Solid Tumors (mRECIST) is a criterion relying on the change of tumor burden before and after treatment [6–8]. Though many image characteristics have been suggested as having prognostic value, substantial subtle features were omitted during traditional imaging assessment, which is highly dependent on individual experience and limited by human eye resolution.
CT texture analysis is a post-processing algorithm that allows further definition of the tumor characteristics beyond the perception of human eyes. By conducting texture analysis, large amounts of texture features are extracted from the pre-treatment images, which can reflect tumor heterogeneity, showing both morphological and cellular diversity [9]. It has been widely applied in many cancer types to predict patients’ outcomes [10–13].
Thus, the primary aim of this study was to create a robust model incorporating texture features derived from contrast-enhanced CT combined with general imaging features as well as clinical information to predict treatment response. Secondary analyses aimed to determine the features that predicted the overall survival (OS) and progress-free survival (PFS) in patients with HCC who received TACE.
This study was approved by the Sun Yat-sen Cancer Centre Institutional Review Board (No. B2021-214-01) with a waiver of written informed consent. All methods were carried out in accordance with relevant guidelines and regulations. From January 2014 to June 2022, data on patients with histological diagnoses of HCC were retrieved from our center's databases. Inclusion criteria were (1) patients with contrast-enhanced CT of the abdomen performed before the initiation of treatment; (2) and who received TACE treatment. Exclusion criteria included (1) patients with a single lesion with a maximal diameter of less than 1 cm or not detectable on CT; (2) disseminated disease within the liver precluding the placement of regions of interest (ROIs); (3) received surgery after TACE; (4) no corresponding laboratory test results; (5) the time interval between CT examination and TACE treatment longer than 14 days; and (6) with other malignancies. Patients’ demographics were recorded, including age, gender, BCLC stage, Child − Pugh class, Eastern Cooperative Oncology Group (ECOG) performance status and complications (diabetes or hypertension). Laboratory test results including platelet (PLT) count, alanine transaminase (ALT), aspartate aminotransferase (AST), total bilirubin (TBIL), international normalized ratio (INR), alkaline phosphatase (ALP), albumin (ALB), C-reactive protein (CRP), Alpha-fetoprotein (AFP), Hepatitis B virus (HBV) and hepatitis C virus (HCV) were collected.
CT examinations were performed using 2 scanners with intravenous contrast media. The volume of the contrast media was determined by multiplying the body weight (in kilograms) by 2 to a maximum of 100 mL. The concentration of the iodinated contrast media used was 350 mg/mL with an injection rate of 2 mL/s. The scanning parameters of the 2 scanners were as follows: (1) The 128-channel CT scanner (Discovery CT750, GE Healthcare, US): field of view, 25 cm; matrix, 512 x 512; tube voltage, 120 kVp; tube current, 200–400 mA; reconstructed thickness, 5 mm; (2) The 128-channel CT scanners (Somatom Definition or Definition AS+, Siemens Healthcare, US): field of view, 35 cm; matrix, 512 x 512; tube voltage, 80–120 kVp; tube current, 248–578 mA; reconstructed thickness, 5 mm. Finally, the arterial phase images of the CT examination were anonymized and assigned a research code for assessment of general imaging features and texture features extraction.
All data were reviewed by 2 board-certified radiologists on a dedicated software (ITK-SNAP, v 3.8.0). The senior radiologist (R1) had more than 10 years of cross-sectional imaging experience, while the junior radiologist (R2) had 5 years of cross-sectional imaging experience. This was designed to test for inter-observer agreement. Only the data from the senior radiologist was used for subsequent feature extraction and model construction.
To begin with, they identified all lesions for each patient in consensus and marked the slice of the largest axial diameter of each lesion. Then, they evaluated general imaging features and drew ROIs separately. Four general imaging features were assessed, including (1) the largest tumor diameter, (2) the number of lesions, (3) the presence or absence of portal vein thrombus, and (4) the presence or absence of ascites. ROIs were drawn by strictly delineating around the margin of the mass with careful inclusion of both solid and cystic components but exclusion of adjacent normal structures. If there were multiple lesions, all the lesions would be given a ROI delineation (Suppl 1).
Texture feature extraction was performed on an open-source Python-based radiomics software (PyRadiomics, v 2.2.0). First, all images are normalized and scaled before textual computation. Then, 5 filters were applied, including Laplacian of Gaussian, wavelet, square, square root, logarithm, and exponential filters [14]. Finally, 7 groups of 1511 texture features were extracted, including (1) first-order statistics, (2) shape-based features, (3) gray-level co-occurrence matrix (GLCM), (4) gray-level-dependent matrix (GLDM), (5) neighboring gray tone difference matrix (NGTDM), (6) gray-level size zone matrix (GLSZM), and (7) gray-level run length matrix (GLRLM). More details about these features were tabulated in Table 1 [15].
Texture feature group |
Description |
---|---|
(1) First-order statistics |
Distribution of grey-level intensities |
(2) Shape-based features |
Description of two- and three- dimensional shape and size |
(3) Gray-level co-occurrence matrix (GLCM) |
The spatial relationship of pixel intensities |
(4) Gray-level-dependent matrix (GLDM) |
Gray level dependencies independent from angles |
(5) Neighboring gray tone difference matrix (NGTDM) |
Difference between gray-level and the average within certain distances |
(6) Gray-level size zone matrix (GLSZM) |
Description of the size of homogeneous zones for each grey-level in 3 dimensions |
(7) Gray-level run length matrix (GLRLM) |
The number of pairs of gray level value and its length of runs |
First, all quantitative features were tested by intraclass correlation coefficient (ICC), and all qualitative features were tested by kappa score. Features with low inter-rater reproducibility (ICC < 0.8 and kappa score < 0.8) were excluded.
Next, patients were dichotomized into progress-free group, including those who achieved complete response (CR), partial response (PR), stable disease (SD), and progress group, including those who exhibited progressive disease (PD) during follow-up. Univariate logistic regression was conducted to select features that had independent prognostic value (p < 0.1).
Finally, the least absolute shrinkage and selection operator (LASSO) algorithm was employed for further feature reduction. The tuning parameter (λ) was selected using 10-fold cross-validation and minimum criteria. A plot of the partial likelihood deviance was made against log (λ). The minimum (lambda.min) and 1-SE criteria (lambda.1se) were used to draw the dotted vertical lines at the optimal values (Fig. 1, Suppl 2).
Patients were given one of the three treatments, conventional TACE (c-TACE) using cytotoxic drugs, drug-eluting beads TACE (DEB-TACE) using chemotherapeutic agents, or microwave ablation with TACE (MWA-TACE), as determined by local multi-disciplinary team in accordance with the recommendations of the European/American Association for Liver Disease guidelines [16, 17].
TACE was performed through femoral access under moderate sedation using the Seldinger technique [18]. To cause embolization of the tumor microcirculation, cytotoxic drugs or chemotherapeutic agents suspended in lipiodol were administrated into the tumor feeding artery with a dose ranging from 5 mL to 30 mL depending on the location, the size, and the number of lesions. If necessary, gelatin sponge particles (150–350 µm) were injected to block the blood until the flow was static.
For patients who received MWA-TACE, CT-guided MWA was performed within 7 days after TACE. One or two 14 G antennae were inserted into the target lesion with the microwave power set from 60 W to 80 W and lasted from 10 to 20 minutes.
All patients were followed up by telephone or clinical visits once every 2 months during the first year and once every 3 months after that until death or the last follow-up day (30th June 2022). Physical examination, hepatic function tests, AFP level, and post-treatment contrast-enhanced CT were reviewed. Their treatment response was evaluated by mRECIST. CR was defined as no intratumorally arterial enhancement in all target lesions. PR was defined as an over 30% reduction of the sum of diameters of target lesions. SD was defined as neither PR nor PD. PD was defined as an over 20% increase in the sum of the diameters of target lesions. OS was defined as the time from baseline CT to death or censoring date. PFS was defined as the time from TACE to disease progression (local recurrence or distant organ metastasis), death, or censoring date.
Data were described as mean and standard deviation or median and range tested by the Shapiro-Wilk test. Fisher’s exact test and Welch’s T-test were used to verify differences among features. Dice coefficient was calculated between the ROIs drawn by the two radiologists. Kappa score and ICC were used to evaluate feature reproducibility between the two radiologists. A random forest classifier was created to differentiate the progress-free group from the progress group. Random survival forest and Cox proportional hazards models were used to evaluate OS and PFS in patients with HCC treated with TACE. A p < 0.05 was considered statistically significant. Statistical analysis was conducted using R software (version 3.5.1).
A total of 301 patients with HCC who received TACE treatment were retrospectively included in this study, with an average age of 55.3 years (range: 22–93 years). Most of them (N = 273, 90.7%) were male. The median time interval from baseline CT examination to TACE treatment was 4.5 days (range: 1–14 days). Overall, 253 patients underwent the cTACE procedure, 18 patients underwent DEB-TACE, and 30 patients underwent MWA-TACE. Patients were randomly allocated into training and testing sets in the ratio of 8:2 for analytical purposes. There was no difference in any clinical features between training and testing sets (p ≥ 0.05). Detailed patient characteristics were reported in Table 2. Detailed results of the assessment of general imaging features were tabulated in Suppl 3.
Whole cohort |
Training set |
Testing set |
p |
||
---|---|---|---|---|---|
N |
301 |
241 |
60 |
||
Age (years) |
55.3 ± 12.4 |
55.9 ± 11.8 |
52.9 ± 14.3 |
0.091 |
|
Gender |
Male |
273 (90.7%) |
219 (90.9%) |
54 (90.0%) |
0.261 |
Female |
28 (9.3%) |
22 (9.1%) |
6 (10.0%) |
||
ECOG performance status |
0 |
298 (99.0%) |
238 (98.8%) |
60 (100.0%) |
0.512 |
1 |
3 (1.0%) |
3 (1.2%) |
0 (0.0%) |
||
Complications |
0 |
256 (85.0%) |
206 (85.5%) |
50 (83.3%) |
0.687 |
1 |
45 (15.0%) |
35 (14.5%) |
10 (16.7%) |
||
Hepatitis |
0 |
48 (15.9%) |
37 (15.4%) |
11 (18.3%) |
0.694 |
1 |
253 (84.1%) |
204 (84.6%) |
49 (81.7%) |
||
ALB (g/L) |
39.2 ± 4.8 |
39.1 ± 4.9 |
39.5 ± 4.7 |
0.526 |
|
ALT (u/L) |
46.3 (9.7-1053.7) |
46.9 (9.7-1053.7) |
43.3 (14.9-249.9) |
0.291 |
|
AST (u/L) |
64.8 (12.5–2690.0) |
65.1 (12.6–2680.0) |
61.0 (12.5-405.9) |
0.508 |
|
TBIL (µmol/L) |
17.2 ± 10.1 |
17.2 ± 10.4 |
17.2 ± 8.6 |
0.962 |
|
PT (s) |
12.2 ± 1.2 |
12.2 ± 1.2 |
12.1 ± 1.2 |
0.553 |
|
INR |
1.1 ± 0.1 |
1.1 ± 0.1 |
1.1 ± 0.1 |
0.550 |
|
PLT (×109/L) |
202.0 (40.0-653.0) |
199.0 (45.0-653.0) |
214.5 (40–649) |
0.233 |
|
Child-Pugh class |
A |
283 (94.0%) |
226 (93.8%) |
57 (95.0%) |
0.778 |
B |
18 (6.0%) |
15 (6.2%) |
3 (5.0%) |
||
CRP (mg/L) |
11.9 (0.2–386.0) |
11.1 (0.2–386.0) |
73.6 (33.5-143.8) |
0.256 |
|
BCLC stage |
A |
64 (21.2%) |
51 (21.2%) |
13 (21.7%) |
0.061 |
B |
95 (31.6%) |
69 (28.6%) |
26 (43.3%) |
||
C |
142 (47.2%) |
121 (50.2%) |
21 (35.0%) |
||
AFP (µg/L) |
773.8 (1.1-865569.0) |
856.6 (1.1-121000.0) |
264.0 (2.0-865569.0) |
0.130 |
|
ECOG 0 = fully active; 1 = restricted in physically strenuous activity | |||||
Complications 0 = without any complications; 1 = with complications (diabetes or hypertension) | |||||
Hepatitis 0 = without hepatitis; 1 = with hepatitis B or with hepatitis C | |||||
Child-Pugh score A = liver is functioning well; B = moderate cirrhosis | |||||
BCLC staging A = early stage, a single tumour of any size, or up to 3 tumors all less than 3 cm; B = multiple tumors in the liver; C = metastasis to the blood vessels, lymph nodes or other body organs |
The mean Dice similarity coefficient of the ROIs between two radiologists was 0.90 ± 0.08. All general imaging features (the diameter of the largest lesion, the number of lesions, the presence or absence of portal vein thrombosis, and the presence or absence of ascites) showed good inter-rater reproducibility (ICC ≥ 0.8 for quantitative features and kappa score ≥ 0.8 for qualitative features). Good inter-rater reproducibility (ICC ≥ 0.8) was shown in 1480 texture features.
First, a total of 1511 texture features were extracted for each patient. The features with low inter-rater reproducibility were excluded, thus reducing the number of features to 1480. Next, univariate logistic regression and LASSO algorithm were conducted to select features that had prognostic value. Finally, 22 features, including 4 clinical features (AFP level, PLT level, ALT level, BCLC stage), 2 general imaging features (the presence or absence of portal vein thrombosis and the presence or absence of ascites) and 16 texture features, were included in further model construction.
During the follow-up, a small proportion of patients (23.6%) achieved progress-free outcome (CR N = 1; PR N = 41 and SD N = 29), while the rest (76.4%) had progressive disease (PD N = 230). The patients were dichotomized into a progress group and progress-free group for analytic purposes. The median PFS and OS were 55 (range: 5-1074) days and 281 (range: 47-2578) days, respectively.
Our random forest model based on the selected 22 features resulted in an AUC of 0.968 with a 95% confidence interval (CI) of 0.932-1.000 for predicting treatment response after TACE. The accuracy was 88.3% (95% CI: 77.4%-95.2%). AFP level, the texture feature square_glszm_LargeAreaEmphasis and PLT count were the top 3 features with the highest importance in the model. The Gini importance was plotted in Fig. 2.
To test the added value of texture features to the predictive model, another random forest model based on the selected 4 clinical features (AFP level, PLT level, ALT level, BCLC stage) and 2 general imaging features (the presence or absence of portal vein thrombosis and the presence or absence of ascites) resulted in an AUC of 0.816 with a 95% confidence interval (CI) of 0.734–0.892 for predicting treatment response after TACE. The accuracy was 69.8% (95% CI: 77.4%-95.2%), significantly inferior to the previous model (p = 0.035).
The mean follow-up period was 36.8 months. On the last follow-up day, 130 patients alive and 43 patients without disease progression were documented as censored for the OS and PFS analyses, respectively. Any kind of death event not limited to HCC-specific related cases were documented in 171 patients. Disease progression was monitored in 258 patients.
The selected 22 features were included in random survival forest and Cox proportional hazards models for predicting OS in HCC treated with TACE. The random forest algorithm based on 180 trees offered the best predictive performance for OS with the lowest OOB Error Rate. Along with the top 3 important clinical and general imaging features (high BCLC stage HR = 1.777, p < 0.001; the presence of portal vein thrombus HR = 1.296, p < 0.001; and high AFP level HR = 1.412, p < 0.001), texture feature wavelet-HLH_glszm_zoneentropy is of the highest importance with a HR of 1.293 (p < 0.001) (Figs. 3 & 4).
The random forest algorithm based on 200 trees offered the best predictive performance for PFS with the lowest OOB Error Rate. Along with the top 3 important clinical imaging features (high AFP level HR = 1.684, p < 0.001; high ALT level HR = 1.296, p = 0.011; and high BCLC stage HR = 1.412, p < 0.001), texture feature wavelet.LLL_firstorder_InterquartileRange is of the highest importance with a HR of 1.337 (p < 0.001) (Figs. 5 & 6).
Our study constructed a random forest classifier that incorporated CT texture features, general imaging features and clinical information in predicting treatment response in HCC after TACE. This model achieved good performance with an AUC of 0.968, offering an objective and non-invasive method for evaluating TACE treatment, potentially avoiding extra imaging examinations or diagnostic work-up and facilitating personalized treatment.
It has been well established that high tumor burden, impaired liver function, incomplete necrosis, the occurrence of extrahepatic spread and vascular invasion should reduce the therapeutic effectiveness of TACE [19]. Traditionally, the treatment decision was made by comprehensive evaluation of physical exams, imaging and laboratory tests, etc. which might be subjective and relies on personal experience. Attempts have been made using texture analysis in HCC as it allows comprehensive and objective characterization of the disease [20]. Our results showed that the selected texture features demonstrated low inter-observer variability regardless of the experience of the radiologists.
A series of studies have been published using texture analysis or radiomic techniques to predict the response of TACE in HCC across different modalities, such as non-contrast CT [21], contrast CT [22], and MRI [23], with good performance (AUC ranging from 0.884–0.960). Tumor heterogeneity and size were identified as critical prognostic features in the Nested multiparametric decision tree [24]. The histogram-based features and shape features were reported to be sensitive in determining the nature of the tumor, which is related to tumoral heterogeneity [25, 26]. Our study achieved a similar or higher AUC compared to the above- mentioned studies and several texture features showed high importance in our classifier, especially for the 3-diension grey level in a matrix classified by region volume, the GLSZM. These features indirectly express a higher degree of tissue homogeneity, which may be interpreted as a consequence of TACE treatment response, thus reducing the contrast between neighboring voxels in non-tumoral component. The CT textural analysis markedly added to the information generated by the clinical parameters in the predictive model.
Prior studies have proven the value of several features in predicting survival in HCC, including AFP level, ALT level, BCLC stage, and the presence or absence of portal vein thrombosis [27–30], similarly those features showed high importance in our random survival forest models. There are two main differences between our study and the prior studies. First, our study extracted several texture features to characterize tumor nature also demonstrated high importance. Second, prior studies constructed models based on radiomics or clinical scores that might have dependencies. Their combination without considering the potential correlation among these dependencies might lead to overfitting the data, in contrast, the random forest algorithms in our study can prevent overfitting by simply reducing tree depth.
There were limitations in our study. First, a small proportion of the patients received DEB-TACE (5.98%) and MWA-TACE (9.97%). The inhomogeneous treatment strategy may affect patients’ prognoses. However, a systematic review revealed that DEB-TACE failed to increase the survival advantage over cTACE [31]. MWA-TACE was reported to show a higher survival rate than those treated with TACE alone in early-stage HCC patients [32], but only a small proportion in our cohort underwent MWA-TACE, and most of the patients in our study were in an intermediate or advanced stage. Second, only arterial phase images were analyzed in our study as previous studies have proven that the extracellular volume and blood flow in HCC during the arterial phase could give rise to unique radiological features [33–35]. Third, selection bias could have resulted from the fact that the patients were recruited from a single specialized oncology medical center, and that by the time they sought treatment here, their disease may already be advanced. Finally, the results from this study were based on texture features extracted using one software. They may not be applicable when using other platforms with different analysis algorithms or higher-order statistics. Standardization and data reproducibility are important before CT texture analysis can be widely applied in clinics.
The current study showed that using random forest algorithms based on the combination of clinical information, general imaging features and texture features derived from pre-treatment contrast-enhanced CT could predict treatment response and survival in HCC treated with TACE. Our findings potentially help patients with HCC avoid additional examinations and assist in treatment planning.
alanine transaminase (ALT)
Albumin (ALB)
alkaline phosphatase (ALP)
Alpha-fetoprotein (AFP)
area under the curve (AUC)
aspartate aminotransferase (AST)
Barcelona Clinic Liver Cancer (BCLC)
Chinese University Prognostic Index (CUPI)
complete response (CR)
computed tomography (CT)
confidence interval (CI)
continuous ranked probability score (CRPS)
conventional TACE (c-TACE)
C-reactive protein (CRP)
drug eluting beads TACE (DEB-TACE)
Eastern Cooperative Oncology Group (ECOG)
gray-level co-occurrence matrix (GLCM)
gray-level run length matrix (GLRLM)
gray-level size zone matrix (GLSZM)
gray-level-dependent matrix (GLDM)
hazard ratio (HR)
hepatitis B virus (HBV)
hepatitis C virus (HCV)
hepatocellular carcinoma (HCC)
Hong Kong Liver Cancer (HKLC)
international normalized ratio (INR)
intraclass correlation coefficient (ICC)
least absolute shrinkage and selection operator (LASSO)
microwave ablation with TACE (MWA-TACE)
modified Response Evaluation Criteria In Solid Tumors (mRECIST)
neighboring gray tone difference matrix (NGTDM)
overall survival (OS)
partial response (PR)
platelet (PLT)
progressive disease (PD)
progress-free survival (PFS)
progressive disease (PD)
receiver operating characteristic curve (ROC)
regions of interest (ROIs)
stable disease (SD)
total bilirubin (TBIL)
transarterial chemoembolization (TACE)
Ethics approval and consent to participate
This study was approved by the Sun Yat-sen Cancer Centre Institutional Review Board (No. B2021-214-01) with a waiver of written informed consent.
Consent for publication
Not applicable.
Availability of data and materials
The anonymized original data is available from the corresponding author, upon reasonable request.
Competing interests
The authors of this manuscript declare no relationships with any companies, whose products or services may be related to the subject matter of the article.
Funding
Not applicable.
Authors' contributions
Conceptualization: HA
Methodology: HA, IB
Software: HA
Validation: HA
Formal analysis: HA
Investigation: HA
Resources: HA, CX
Data Curation: HA, IB
Writing - Original Draft: HA
Writing - Review & Editing: HA, IB, CX
Visualization: HA
Supervision: CX
Project administration: CX
Funding acquisition: N.A.
Acknowledgements
Not applicable.