Contributed equally
The purpose of the present study was to investigate the value of contrast-enhanced magnetic resonance imaging (CE-MRI) texture analysis for preoperatively predicting microvascular invasion (MVI) in hepatocellular carcinoma (HCC). Accordingly, a retrospective study of 142 patients with pathologically confirmed HCC was performed. The patients were divided into two cohorts: The training cohort (n=99) and the validation cohort (n=43), including the MVI-positive group (n=53) and MVI-negative group (n=89). On the basis of three-dimensional texture analysis, 58 features were extracted from the preoperative CE-MR images of arterial-phase (AP) and portal-venous-phase (PP). The t-test or Kruskal-Wallis test, univariate logistic regression analysis and Pearson correlation were applied for feature reduction. Clinical-radiological features were also analyzed. Multivariate logistic regression analysis was used to build the texture model and combined model with clinical-radiological features. The MVI-predictive performance of the models was evaluated using receiver operating characteristic (ROC) analysis and presented using nomogram. Among the clinical features, a significant difference was found in maximum tumor diameter (P=0.002), tumor differentiation (P=0.026) and α-fetoprotein level (P=0.025) between the two groups in the training cohort. Four MR texture features in AP and five in PP images were identified through feature reduction. On ROC analysis, the AP texture model showed better diagnostic performance than did the PP model in the validation cohort, with an area under the curve (AUC) of 0.773 vs. 0.623, sensitivity of 0.750 vs. 0.500, and specificity of 0.815 vs. 0.926. Together with the clinical features, the combined model of AP improved the AUC, sensitivity and specificity to 0.810, 0.811 and 0.790, respectively, which was demonstrated in nomogram. To conclude, model-based texture analysis of CE-MRI could predict MVI in HCC preoperatively and noninvasively, and the AP image shows better predictive efficiency than PP image. The combined model of AP with clinical-radiological features could improve MVI prediction ability.
Hepatocellular carcinoma (HCC) is the sixth most common malignant tumor and the second leading cause of tumor mortality worldwide (
In recent years, many clinical studies have shown that microvascular invasion (MVI) is a significant risk factor for the high rate of recurrence and poor prognosis, and could provide information on which to base clinical treatment (
Recently, a number of studies reported that certain clinical features and morphological characteristics detected on imaging examination could predict the MVI status, including the tumor size, tumor margin, capsule formation and dynamic enhancing pattern (
Texture analysis is a widely used image post-processing technique that extracts quantitative features from radiological images to explore the correlation between these features and clinical or histological factors (
In particular, contrast-enhanced MRI (CE-MRI) is generally used for the diagnosis, treatment evaluation, prognosis estimation, as well as MVI prediction of HCC (
The purpose of the present study was to explore the value of CE-MRI texture analysis in preoperatively predicting the MVI status of HCC and determining the diagnostic performance to guide the clinician in choosing appropriate treatment options.
The Independent Ethics Committee of the Cancer Hospital, Chinese Academy of Medical Sciences (Beijing, China) approved the current retrospective study and waived the requirement for informed patient consent. Between January 2014 and December 2016, a total of 282 patients with HCC underwent liver MRI examination confirmed by postoperative pathological examinations. The inclusion criteria were as follows: i) Single tumor with a maximum diameter <5.0 cm, no large vessel invasion and no distant metastasis; ii) underwent radical resection; iii) primary HCC and MVI status confirmed by surgical pathological examination; iv) no other antitumor treatment received before MRI examination and operation; and v) no apparent artifact that may affect imaging analysis.
The pathological diagnostic criteria for MVI adopted in this study were reported by Rodríguez-Perálvarez
The patients fasted for 6–8 h to empty the gastrointestinal tract before undergoing MRI examination. All of the MR images were acquired using a 3.0-T body MRI system (Discovery MR750 3.0T, GE Medical Systems) equipped with an 8-channel phased-array body coil. The CE-MRI acquisitions were performed with multiphase 3D spoiled gradient echo liver acceleration volume acquisition (LAVA) sequence, with the following scanning parameters: Repetition time, 2.9 msec; echo time, 1.3 msec; flip angle, 12°; field of view, 36–42×36–42 cm; matrix, 512×512; section thickness, 4 mm; gap, 0 mm; and number of sections, 36–40. Gadodiamide (Omniscan 0.5 mmol/ml; GE Healthcare) at a standard dose (0.2 ml/kg) was injected as a bolus through the peripheral veins by using an automatic pump injector at the rate of 3.0 ml/sec, and followed immediately by 20 ml of a 0.9% sterile saline solution injection. The contrast-enhanced dynamic images were acquired at 15–20 sec (arterial phase, AP), 50–55 sec (portal venous phase, PP) and 85–90 sec (delayed phase) after contrast-agent injection by using the LAVA sequence.
The basic MRI features included the maximum tumor diameter (MTD) (measured by the maximum diameter on the maximum axial section in PP MR image), liver background (cirrhosis or noncirrhosis), tumor encapsulation (peripheral rim of smooth hyperenhancement in PP image), fast wash-in (hyperenhancement of the tumor in the AP), fast wash-out (hypoenhancement of the tumor in the PP) and tumor necrosis (unenhanced areas). The features for each patient were independently evaluated and recorded in a blinded manner by two radiologists with 5 (YJZ) and 3 years (BF) of experience in the interpretation of abdominal MRI to ensure diagnostic accuracy. When a disagreement occurred between the two reviewers during evaluation, a joint review was performed, and consensus data were used for further statistical analysis.
All the MR images were retrieved from the picture archiving and communication system and transferred to a personal computer in the Digital Imaging and Communications in Medicine format. The same two radiologists reviewed and processed the images in a random patient order by using an in-house developed software, Omni-kinetics (version 2.0.10; GE Healthcare Life Sciences), to obtain texture features. A 3D volume of interest (VOI) of the tumor was manually contoured by the two readers, slightly along the borders of the tumor to include the entire approximated tumor volume.
After generating the VOI, a total of 58 texture features were automatically extracted from the AP and PP images using the Omni-kinetics software. The texture features could be divided into four categories: i) 29 histogram features, ii) 8 gray-level co-occurrence matrix (GLCM) features, iii) 11 Haralick features, and iv) 10 run-length matrix (RLM) features. A detailed list of the features included in the present study is presented in
The Kolmogorov-Smirnov test was used to determine whether the distribution of all the features was normal, and Levene's test was used for identifying the homogeneity of variance. For the clinical-radiological features, a two-tailed unpaired independent t-test was used to compare continuous variables with normal distribution between the MP and MN groups. Categorical variables were compared using the χ2 test or Fisher's exact test.
For texture features, first, an independent t-test or Kruskal-Wallis test was applied one by one. Features with significant differences (P<0.05) were further analyzed by univariate logistic regression analysis. Features in the univariate logistic regression analysis with P<0.05 were selected. Finally, to eliminate redundant features, Pearson correlation analysis was conducted to remove features with high correlation (r>0.90), which were not considered in the subsequent analysis. Features that remained after adjusting for redundancy were entered into model building.
All statistical analyses were performed using R software (version 3.4.1; R Foundation for Statistical Computing), with a two-tailed probability value. P<0.05 was considered to indicate a statistically significant difference.
Multivariate logistic regression analysis was applied for model building. First, the texture model was built on the basis of the features selected from a previous step by directly entering. The texture signature score (Texscore), which reflected the overall texture features, was calculated for each patient using the texture model. To involve both texture features and clinical-radiological features to improve performance, the combined model was built on the basis of the Texscore as well as other significant clinical-radiological features in both AP and PP.
The performance of the texture and combined models was analyzed both in the training and validation cohorts by using a ROC curve quantified by the area under the curve (AUC), sensitivity, specificity and overall accuracy (ACC). The cutoff point was calculated at the maximized value of the Youden index (sensitivity + specificity-1) (
Of the 142 patients enrolled in the study, 124 patients were male and 18 were female. Median age was 57 years (range, 34–80 years). The results revealed that MTD (P=0.002), serum AFP level (P=0.025) and tumor differentiation (P=0.026) showed significant differences between the MP and MN groups in the training cohort. The MP group had greater MTD than did the MN group (3.82±0.88 vs. 3.21±0.94 cm), and also tended to have a higher serum AFP level and lower tumor differentiation. This result was also confirmed in the validation cohort. The characteristics of the training and validation cohorts showed no significant differences (all P>0.05). The detailed clinical and radiological features of the patients in the training and validation cohorts are listed in
First, the Kruskal-Wallis test revealed that 19 texture features in AP and 18 in PP showed significant differences (all P<0.05) between the MP and MN groups. In the univariate logistic regression analysis, 10 texture features in AP and 12 in PP showed a potential predictive value (P<0.05) to discriminate between the MP and MN groups. Detailed results of the univariate logistic regression analysis are demonstrated in
Texture features identified from the above analysis in AP and PP were entered into the multivariate logistic regression model to build the texture predictive model. The detailed results of the multivariate logistic regression analysis are shown in
The texture signature score (Texscore) of each patient in AP and PP could be calculated using the formula based on multivariate logistic regression model as follows: Texscore (AP) = 0.455+1.30 × uniformity + 0.524 × ClusterProminence + 0.593 × ClusterShade + 0.494 × LRLGLE; and Texscore (PP) = 7.310 + 0.732 × GlcmEntropy + 0.002 × sumEntropy + 2.230×105 × sumAverage + 0.999 × MinIntensity + 1.000 × RLN.
In the training cohort, two combined models were built to predict the MVI status by clinical-radiological features and the Texscore generated above in AP and PP separately (
The nomograms integrating the features included in the multivariate logistic regression analysis are displayed in
For the training cohort, the ROC curves illustrating the predictive performance of the texture and combined models in predicting the MVI status are provided in
The validation cohort was used to verify the accuracy of the model built in the training cohort. The ROC curves of the texture and combined models for predicting the MVI status are shown in
In the present study, a combined model was developed based on preoperative 3D CE-MRI texture and clinical-radiological features to predict MVI with a satisfactory discriminatory performance. The results indicated that texture analysis is a potentially useful adjunct for predicting MVI, and adding clinical-radiological data could slightly improve the predictive ability. In addition, the AP texture features performed better than the PP texture features in MVI prediction.
The present study involved a total of 142 patients, 53 of which were MVI-positive patients and 89 MVI-negative. The MVI-positive rate was consistent with the reported rate in the literature (
CE-MRI texture analysis of the AP and PP images was used to build models for predicting the MVI status, and the AP images showed better predictive ability than did the PP images. Four texture features in AP (uniformity, ClusterProminence, ClusterShade and LRLGLE) and five in PP (MinIntensity, GlcmEntropy, sumAverage, sumEntropy and RLN) were entered into the multivariate logistic regression analysis. These features frequently appeared in texture or radiomics research and showed noteworthy diagnostic and predictive efficiency, which could be explained by their definitions (
In the multivariate logistic regression analysis, the directly enter mode was used to build the predictive model, but the P-values of the texture features in the texture model were not significant. The possible explanation was that all parameters were equally important, and hence, no significant texture features were observed in the regression. The combined model was also built by adding the clinical-radiological features to the multivariate logistic regression. The nomogram was used to visualize the combined model and to reveal the weight of the clinical-radiological and texture features. For clinical use of the model, the total scores of each patient could be calculated based on the nomogram. High scores corresponded to a high probability of MVI occurrence. The nomogram showed that the texture signature accounted for a higher proportion in the total points than did the clinical-radiological features in AP.
In the ROC analysis, the combined model showed a little better predictive performance than did the texture model in the validation cohort both in AP (AUC, 0.794 vs. 0.773; specificity, 0.852 vs. 0.815; sensitivity, 0.812 vs. 0.750) and in PP (AUC, 0.706 vs. 0.623; specificity, 0.704 vs. 0.926; sensitivity, 0.750 vs. 0.500). The predictive ability was better in AP than in PP. The possible explanations might be that the blood supply to the HCC is mainly dependent on the hepatic artery, and that the AP image could more clearly reflect the changes in small blood vessels in the HCC. Moreover, when MVI occurs, the local hemodynamics of the liver tissue around the tumor changes. Therefore, the changes in MRI texture features in AP were more obvious than those in PP.
From a clinical perspective, the results of the present study suggest that CE-MRI texture analysis may be an option for preoperatively predicting MVI in HCC. This could alert pathologists to conduct more detailed pathological examinations, particularly when preoperative texture analysis suggests a high possibility of MVI occurrence. Meanwhile, predicting the possibility of MVI occurrence could also help clinicians select more suitable surgical procedures for HCC patients. A number of studies have shown that anatomic resection (
The present study had the following advantages. First, to our knowledge, this was the first study to use 3D MRI texture features to build a model for predicting the MVI status preoperatively and noninvasively. The 3D VOI may have a better predictive performance since MVI could occur in every slice, and 3D feature analysis considers all of the available slices with abundant information. Previous studies have also shown that 3D features improved the diagnostic accuracy than did two-dimensional features (
The current study also had several limitations. First, this was a retrospective study with a single-center design; therefore, selection bias was unavoidable. Second, the sample size was small for texture analysis. More cases are needed to verify the results. Third, the manual segmentation of tumors may have introduced a certain amount of subjectivity. The MVI grade in the positive group was not taken into account. Forth, the present study did not include des-γ-carboxy-prothrombin (DCP) as it was not tested routinely in our hospital, although a study has reported that DCP was useful for prediction of MVI (
In conclusion, model-based texture analysis of CE-MRI could predict MVI in HCC preoperatively and noninvasively. The AP image shows better predictive efficiency than PP image. The combined model of AP with clinical-radiological features could improve MVI prediction ability. It may be of value to clinicians in objectively selecting appropriate treatment strategies and as an individualized predictive tool for improving clinical outcomes.
Not applicable.
This study was supported by the Peking Union Medical College Youth Fund and the Fundamental Research Funds for the Central Universities (grant no. 2017320010), the Chinese Academy of Medical Sciences (CAMS) Research Fund (grant no. ZZ2016B01) and the CAMS Innovation Fund for Medical Sciences (grant no. 2016-I2M-1-001).
The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.
YJZ, SW, LMW, XHM and XMZ designed this study. YJZ, BF and LMW acquired the patients and searched the database. YJZ and BF participated in the process of texture analysis, including segmentation and features extraction. JFW contributed to the analysis and interpretation of data, including statistical analysis, biostatistics and computational analysis. YJZ and XHM were major contributors in writing the manuscript. All authors read and approved the final manuscript.
The independent ethics committee of the Cancer Hospital, Chinese Academy of Medical Sciences approved this retrospective study and waived the requirement for informed patient consent.
Not applicable.
The authors declare that they have no competing interests.
contrast-enhanced magnetic resonance imaging
microvascular invasion
hepatocellular carcinoma
three-dimensional
arterial phase
portal-venous phase
receiver operating characteristic
area under the curve
surgical margin
positron emission tomography-computed tomography
MVI-positive
MVI-negative
α-fetoprotein
liver acceleration volume acquisition
maximum tumor diameter
volume of interest
gray-level co-occurrence matrix
texture signature score
overall accuracy
LongRunLowGreyLevelEmphasis
RunLengthNonuniformity
confidence interval
Diagram showing the recruitment of the study population and exclusion criteria. HCC, hepatocellular carcinoma; MRI, magnetic resonance imaging; MVI, microvascular invasion.
Schematic diagram showing the MVI diagnosis based on pathological examination, prediction using texture analysis and their relationship. (A) Tumor cells could be found in the small vessels near the primary tumor lesion (black arrow). (B) The process of texture analysis includes image segmentation, feature extraction, statistical analysis and model building. MVI, microvascular invasion; GLCM, gray-level co-occurrence matrix; RLM, run-length matrix.
Pearson correlation matrix plot for (A) 10 texture features in the arterial phase and (B) 12 texture features in the portal venous phase of the training cohort. Blue circles indicate positive correlation, red circles negative correlation. The larger the circle and the darker the color, the higher is the correlation between two texture features. GLN, GreyLevelNonuniformity; LRLGLE, LongRunLowGreyLevelEmphasis; LGLRE, LowGreyLevelRunEmphasis; SRLGLE, ShortRunLowGreyLevelEmphasis; RLN, RunLengthNonuniformity; ASM, AngularSecondMoment; UPP, uniformity of distribution of positive gray-level pixel values.
Nomograms for predicting the microvascular invasion status of hepatocellular carcinoma in (A) the arterial phase and (B) portal venous phase using the texture signature and clinical-radiological features. AFP) levels 1, 2 and 3 stand for ≤7, 7–400 and >400 ng/ml, respectively. Differentiation levels 1, 2 and 3 stand for poor, moderate and well differentiation, respectively. MTD, maximum tumor diameter; AFP, α-fetoprotein.
ROC curves for the texture and combined models in predicting the microvascular invasion status in the training cohort. The solid dots represent the optimal cutoff values (specificity, sensitivity) for discrimination. The AUC for the texture models in (A) AP and (B) PP are 0.765 and 0.707, respectively. The AUC for the combined models in (C) AP and (D) PP are 0.810 and 0.799, respectively. ROC, receiver operating characteristic; AUC, area under the curve; AP, arterial phase; PP, portal-venous phase.
ROC curves for the texture and combined models in predicting the microvascular invasion status in the validation cohort. The solid dots represent the optimal cutoff values (specificity, sensitivity) for discrimination. The AUC for the texture models in the (A) AP and (B) PP are 0.773 and 0.623, respectively. The AUC for the combined models in (C) AP and (D) PP are 0.794 and 0.706, respectively. ROC, receiver operating characteristic; AUC, area under the curve; AP, arterial phase; PP, portal-venous phase.
List of 58 texture analysis parameters.
Texture type | Texture parameters |
---|---|
Histogram | MinIntensity, MaxIntensity, MedianIntensity, MeanValue, stdDeviation, Variance, VolumeCount, VoxelValueSum, RMS, Range, MeanDeviation, RelativeDeviation, MinLocation, MaxLocation, Skewness, kurtosis, uniformity, Energy, Entropy, FrequencySize, MPP, UPP, Quantile5, Quantile10, Quantile25, Quantile50, Quantile75, Quantile90, Quantile95 |
GLCM | GlcmTotalFrequency, GlcmEnergy, GlcmEntropy, Inertia, Correlation, InverseDifferenceMoment, ClusterShade, ClusterProminence |
Haralick | HaralickCorrelation, HaraEntropy, AngularSecondMoment, contrast, HaraVariance, sumAverage, sumVariance, sumEntropy, differenceVariance, differenceEntropy, inverseDifferenceMoment |
RLM | ShortRunEmphasis, LongRunEmphasis, GreyLevelNonuniformity, RunLengthNonuniformity, LowGreyLevelRunEmphasis, HighGreyLevelRunEmphasis, ShortRunLowGreyLevelEmphasis, ShortRunHighGreyLevelEmphasis, LongRunLowGreyLevelEmphasis, LongRunHighGreyLevelEmphasis |
RMS, Root mean square; GLCM, gray-level co-occurrence matrix; MPP, mean value of positive pixels; UPP, uniformity of distribution of positive gray-level pixel values; RLM, run-length matrix.
The clinical and radiological characteristics of patients in the training and validation cohorts.
A, Clinical characteristics | |||||||
---|---|---|---|---|---|---|---|
Training cohort (n=99) | Validation cohort (n=43) | ||||||
Characteristic | MP (n=37) | MN (n=62) | P-value |
MP (n=16) | MN (n=27) | P-value |
P-value |
Sex | 0.931 |
0.723 |
0.805 |
||||
Male | 32 | 54 | 15 | 23 | |||
Female | 5 | 8 | 1 | 4 | |||
Mean age ± SD, years | 57.49±9.56 | 56.45±9.71 | 0.607 |
55.88±8.63 | 55.41±8.00 | 0.858 |
0.456 |
AFP, ng/ml | 0.025 |
0.038 |
0.715 |
||||
≤7 | 9 | 32 | 5 | 16 | |||
7–400 | 17 | 20 | 5 | 9 | |||
>400 | 11 | 10 | 6 | 2 | |||
Location | 0.561 |
0.372 |
0.581 |
||||
Right | 31 | 49 | 10 | 23 | |||
Left | 6 | 13 | 6 | 4 | |||
HBsAg | 0.325 |
0.614 |
0.668 |
||||
Positive | 26 | 49 | 12 | 22 | |||
Negative | 11 | 13 | 4 | 5 | |||
HCV-Ab | 1.000 |
1.000 |
0.989 |
||||
Positive | 1 | 3 | 0 | 1 | |||
Negative | 36 | 59 | 16 | 26 | |||
Differentiation | 0.026 |
0.034 |
0.565 |
||||
Well | 2 | 3 | 0 | 4 | |||
Moderate | 19 | 47 | 9 | 20 | |||
Poor | 16 | 12 | 7 | 3 | |||
MTD, cm | 3.82±0.88 | 3.21±0.94 | 0.002 |
3.75±0.80 | 2.92±0.62 | <0.001 |
0.231 |
Background liver | 0.690 |
0.358 |
0.553 |
||||
Noncirrhosis | 14 | 21 | 3 | 10 | |||
Cirrhosis | 23 | 41 | 13 | 17 | |||
Tumor encapsulation | 0.805 |
0.362 |
0.534 |
||||
Absent | 11 | 17 | 2 | 8 | |||
Present | 26 | 45 | 14 | 19 | |||
Fast wash-in | 0.292 |
0.929 |
0.742 |
||||
Yes | 30 | 55 | 14 | 22 | |||
No | 7 | 7 | 2 | 5 | |||
Fast wash-out | 0.409 |
0.534 |
0.720 |
||||
Yes | 24 | 35 | 11 | 16 | |||
No | 13 | 27 | 5 | 11 | |||
Tumor necrosis | 0.638 |
0.372 |
0.668 |
||||
Absent | 29 | 46 | 11 | 23 | |||
Present | 8 | 16 | 5 | 4 |
MP, microvascular invasion-positive; MN, microvascular invasion-negative; AFP, α-fetoprotein; HBsAg, hepatitis B surface antigen; MTD, maximum tumor diameter; HCV-Ab, hepatitis C antibody.
MP vs. MN
Training cohort vs. validation cohort
P-values calculated using the χ2 test
P-values calculated using Fisher's exact test
P-values calculated using the independent t-test.
Univariate logistic regression analysis of the clinical and texture features to predict the microvascular invasion status in the training cohort.
OR | ||||
---|---|---|---|---|
95% CI | ||||
Feature | Value | Lower | Upper | P-value |
AP texture features | ||||
Uniformity | 1.209×104 | 6.602 | 4.605×107 | 0.019 |
Energy | 3.110×10121 | 4.556×1019 | 1.820×10229 | 0.022 |
Entropy | 1.123×10−1 | 1.991×10−2 | 5.647×10−1 | 0.010 |
UPP | 3.110×10121 | 4.556×1019 | 1.820×10229 | 0.022 |
ClusterShade | 1.000 | 1.000 | 1.000 | 0.046 |
ClusterProminence | 1.000 | 1.000 | 1.000 | 0.005 |
GreyLevelNonuniformity | 1.003 | 1.001 | 1.007 | 0.023 |
LowGreyLevelRunEmphasis | 0.000 | 0.000 | 3.320 ×10−192 | 0.034 |
ShortRunLowGreyLevelEmphasis | 0.000 | 0.000 | 2.140×10−255 | 0.035 |
LongRunLowGreyLevelEmphasis | 0.000 | 0.000 | 2.060 ×10−79 | 0.036 |
PP texture features | ||||
MinIntensity | 9.973×10−1 | 9.947×10−1 | 9.995×10−1 | 0.025 |
VolumeCount | 1.000 | 1.000 | 1.000 | 0.027 |
Uniformity | 1.087×10−3 | 9.070×10−7 | 6.241×10−1 | 0.044 |
FrequencySize | 1.000 | 1.000 | 1.000 | 0.027 |
GlcmTotalFrequency | 1.000 | 1.000 | 1.000 | 0.026 |
GlcmEntropy | 6.627×10−1 | 4.454×10−1 | 9.462×10−1 | 0.030 |
HaraEntroy | 3.266×105 | 2.273×101 | 1.286×1010 | 0.013 |
AngularSecondMoment | 0.000 | 0.000 | 4.329×10−94 | 0.026 |
sumAverage | 1.301×10−2 | 2.238×10−4 | 5.748×10−1 | 0.029 |
sumEntropy | 1.896×104 | 4.127 | 1.996×108 | 0.028 |
GreyLevelNonuniformity | 1.003 | 1.001 | 1.006 | 0.024 |
RunLengthNonuniformity | 1.000 | 1.000 | 1.000 | 0.025 |
OR, odds ratio; CI, confidence interval; UPP, uniformity of distribution of positive gray-level pixel values; AP, arterial phase; PP, portal venous phase.
Multivariate logistic regression analysis of the texture parameters in predicting the microvascular invasion status in the arterial and portal venous phases in the training cohort.
A, Arterial phase | ||
Feature | OR | P-value |
---|---|---|
(Intercept) |
0.455 | 0.003 |
Uniformity | 1.301 | 0.387 |
ClusterShade | 0.524 | 0.059 |
ClusterProminence | 0.593 | 0.262 |
LRLGLE | 0.494 | 0.170 |
Feature | OR | P-value |
(Intercept) |
7.310 | 0.754 |
GlcmEntropy | 0.732 | 0.146 |
sumEntropy | 0.002 | 0.474 |
sumAverage | 2.230e+05 | 0.124 |
MinIntensity | 0.999 | 0.563 |
RLN | 1.000 | 0.371 |
Intercept is the constant term in the logistic regression equation. OR, odds ratio; LRLGLE, LongRunLowGreyLevelEmphasis; RLN, RunLengthNonuniformity.
Multivariate logistic regression analysis of the combined clinical-radiological and texture features to predict the microvascular invasion status in the AP and PP in the training cohort.
AP Model | PP Model | |||
---|---|---|---|---|
Features | OR | P-value | OR | P-value |
(Intercept) |
0.257 | 0.282 | −3.068 | 0.052 |
MTD, cm | 1.395 | 0.242 | 0.759 | 0.032 |
Differentiation | ||||
Moderate vs. poor | 0.559 | 0.284 | −0.838 | 0.179 |
Well vs. low | 0.672 | 0.747 | −0.697 | 0.560 |
α-fetoprotein, ng/ml | ||||
7–400 vs. ≤7 | 1.801 | 0.298 | 1.307 | 0.027 |
>400 vs. ≤7 | 3.771 | 0.053 | 1.768 | 0.017 |
Texscore | 2.552 | 0.003 | 0.449 | 0.294 |
Intercept is the constant term in the logistic regression equation, and is not a clinical-radiological feature. AP, arterial phase; PP, portal venous phase; OR, odds ratio; MTD, maximum tumor diameter.
Predictive performance of the texture and combined models in predicting the microvascular invasion status in the training and validation cohorts.
Training cohort (n=99) | Validation cohort (n=43) | ||||||||
---|---|---|---|---|---|---|---|---|---|
Model | Phase | AUC | ACC | SEN | SPE | AUC | ACC | SEN | SPE |
Texture | AP | 0.765 | 0.768 | 0.730 | 0.790 | 0.773 | 0.791 | 0.750 | 0.815 |
PP | 0.707 | 0.727 | 0.622 | 0.790 | 0.623 | 0.767 | 0.500 | 0.926 | |
Combined | AP | 0.810 | 0.798 | 0.811 | 0.790 | 0.794 | 0.837 | 0.812 | 0.852 |
PP | 0.799 | 0.758 | 0.730 | 0.774 | 0.706 | 0.721 | 0.750 | 0.704 |
AP, arterial phase; PP, portal venous phase; AUC, area under the curve; ACC, overall accuracy; SEN, sensitivity; SPE, specificity.