Abstract
Purpose:
The clinical adoption of quantitative imaging biomarkers (radiomics) has established the need for high quality contrast-enhancement in medical images. We aimed to develop a machine-learning algorithm for Quality Control of Contrast-Enhancement on CT-scan (CECT-QC).
Method:
Multicenter data from four independent cohorts [A, B, C, D] of patients with measurable liver lesions were analyzed retrospectively (patients:time-points; 503:3397): [A] dynamic CTs from primary liver cancer (60:2359); [B] triphasic CTs from primary liver cancer (31:93); [C] triphasic CTs from hepatocellular carcinoma (121:363); [D] portal venous phase CTs of liver metastasis from colorectal cancer (291:582). Patients from cohort A were randomized to training-set (48:1884) and test-set (12:475). A random forest classifier was trained and tested to identify five contrast-enhancement phases. The input was the mean intensity of the abdominal aorta and the portal vein measured on a single abdominal CT scan image at a single time-point. The output to be predicted was: non-contrast [NCP], early-arterial [E-AP], optimal-arterial [O-AP], optimal-portal [O-PVP], and late-portal [L-PVP]. Clinical utility was assessed in cohorts B, C, and D.
Results:
The CECT-QC algorithm showed performances of 98 %, 90 %, and 84 % for predicting NCP, O-AP, and O-PVP, respectively. O-PVP was reached in half of patients and was associated with a peak in liver malignancy density. Contrast-enhancement quality significantly influenced radiomics features deciphering the phenotype of liver neoplasms.
Conclusions:
A single CT-image can be used to differentiate five contrast-enhancement phases for radiomics-based precision medicine in the most common liver neoplasms occurring in patients with or without liver cirrhosis.
Keywords: Machine learning, Radiomics, Quality control, Contrast media, Liver neoplasms
1. Introduction
Recent progress in the fields of radiomics and artificial intelligence (AI) has transformed medical imaging by enabling high-throughput data mining and automated analysis of patterns present in images from standard-of-care CT scans. Since precision medicine strategies are now integrating imaging biomarkers, a major challenge is to ensure that the quality of imaging datasets is sufficient to provide robust, reproducible, and clinically meaningful biomarker panels [1].
AI techniques developed based on radiographic images require rigorous quality control, standardized collection, and large-scale sharing of data. A recent study demonstrated that the contrast-enhancement quality of a CT-scan acquired at the portal venous phase in multicenter clinical trials was not optimal in one third of colorectal cancer patients [2]. Therefore, we aimed to develop a novel Contrast-Enhancement CT-scan Quality Control (CECT-QC) algorithm to provide improved standardization and quality control in the evaluation of contrast-enhancement quality. Using machine-learning, the CECT-QC algorithm was trained to identify the optimal contrast-enhancement phases of abdominal CT scan images.
We aimed to prove that such a CECT-QC algorithm could be a valuable tool to guide decision-support systems. Consequently, we applied the CECT-QC algorithm to critical clinical indications that could leverage radiomics-based precision diagnoses and treatments.
The CECT-QC algorithm was developed in patients with hepatocellular carcinoma (HCC). HCC is the most common primary hepatic malignancy and the fourth leading cause of cancer-related deaths worldwide [[3] [4],]. The CECT-QC algorithm was investigated using cirrhotic patients as the target patient population since cirrhosis, the main risk factor for developing HCC, alters the biodistribution of contrast-enhancement product in all tissues. Pathophysiology of cirrhosis as well as cirrhotic cardiomyopathy can significantly alter the extraction of imaging biomarkers in HCC tissues by altering the biodistribution of the contrast-enhancement product and delay or decrease portal vein enhancement [5–7]. Therefore, optimal liver contrast-enhancement is challenging in cirrhotic patients.
The diagnosis of HCC could benefit from radiomic analyses of abdominal CECT [8]. HCC treatment can be initiated based on non-invasive imaging diagnostic criteria alone in most patients, unlike other solid tumors which require biopsy confirmation [9]. However, the guidelines by the European Association for the Study of the Liver (EASL) define indeterminate liver nodules as liver nodules in cirrhotic patients that do not present the distinctive patterns of HCC, and therefore can’t be diagnosed non-invasively [10–12]. Radiologists’ difficulty in evaluating these indeterminate liver nodules on cross-sectional images may be mitigated in part by radiomic evaluation. In cirrhotic patients, AI could be trained to detect and provide a diagnosis for otherwise indeterminate liver lesions and could therefore preclude the morbidity and mortality associated with current alternative diagnostic strategies such as liver biopsy, or a wait-and-see strategy [13,14].
Response assessment in non-resectable liver malignancies is another problem that could benefit from radiomic-guided precision treatment optimizing the balance between treatment benefit and risk, as well as cost-effectiveness [15]. AI could be trained to predict antitumor activity of systemic therapies by deciphering spatial and temporal changes in tumor imaging phenotypes [16]. An example of an imaging biomarker is tumor density measured on a CT-scan. Tumor density is a surrogate for tumor vascularity [17–19] and has the potential for wide-ranging applications across cancer diagnoses and therapies. This imaging biomarker has been validated by several international criteria [18–21].
The aim of this study was to develop and validate a contrast-enhancement phase-recognition algorithm: the quality control of contrast-enhancement of CT-scan images (CECT-QC algorithm). A random forest classifier was trained and tested to identify five contrast-enhancement phases. The input was the mean intensity of the abdominal aorta and the portal vein measured on a single abdominal CT scan image at a single time-point. The output to be predicted was five contrast-enhancement phases. This algorithm was then applied to external independent clinical datasets to evaluate its usefulness in the previously defined indications.
2. Material and methods
2.1. Study overview
Multicenter data from four independent cohorts [A, B, C, D] totaling 503 patients with measurable liver lesions evaluated at 3397 time-points were analyzed retrospectively: [A] dynamic CTs from 60 patients with primary liver cancer (2359 time-points); [B] triphasic CTs from 31 cirrhotic patients with primary liver cancer (93 time-points); [C] triphasic CTs from 121 cirrhotic patients with hepatocellular carcinoma (363 time-points); [D] 291 patients with liver metastasis from colorectal cancer at the PVP (2 acquisitions for each patient, 582 time-points). Patients from cohort A, were randomized to training-set and test-set (48:12 patients, 1884:475 time-points).
Institutional review board approved this retrospective HIPAA-compliant study and all CT-scans were performed as standard of care.
In the first part of this study, we trained and tested a phase classifier, i.e., the CECT-QC algorithm, to differentiate five contrast-enhancement phases in cohort A.
In the second part of this study, we applied the CECT-QC algorithm to three independent clinical cohorts: B, C, D. For this purpose, we first evaluated the percentage of optimal acquisitions in these three datasets. Then, we estimated how non-optimal CE altered the measurement of three types of imaging biomarkers: (i) tumor density, (ii) ‘washout’ in tumor density, and (iii) radiomic ‘washout’ features. Details of the four datasets are summarized in Fig. 1 and Table 1.
Fig. 1.
Flow chart.
Table 1.
Patient characteristics.
| Cohort A | Cohort B | Cohort C | Cohort D | |
|---|---|---|---|---|
| Training set | Yes | No | No | No |
| Cirrhosis | Yes | Yes | Yes | No |
| CT-scan acquisition | Dynamic CTs | Triphasic CTs | Triphasic CTs | PVP CTs |
| Histology | Liver nodule | Non-HCC | HCC | Colorectal cancer |
| Number of patients | 60 | 31 | 121 | 291 |
| Age (years) | 52 ± 13 | 61 ± 12 | 66 ± 10 | / |
| Male | 50 (0.83) | 25 (0.81) | 105 (0.87) | 186 (0.64) |
| Female | 10 (0.17) | 6 (0.19) | 16 (0.13) | 105 (0.36) |
| Cirrhosis cause* | ||||
| Alcohol | 8 (.13) | 14 (.45) | 56 (.46) | / |
| Hepatitis B | 45 (.75) | 1 (.3) | 7 (.6) | / |
| Hepatitis C | 0 (.0) | 4 (.13) | 29 (.24) | / |
| HIV | 0 (.0) | 1 (.3) | 1 (.1) | / |
| Hemochromatosis | 0 (.0) | 0 (.0) | 3 (.2) | / |
| NASH | 0 (.0) | 3 (.10) | 7 (.6) | / |
| Unreported | 15 (.25) | 8 (.26) | 18 (.15) | / |
| CHILD score at diagnosis | / | |||
| A | 0 (.0) | 14 (.45) | 61 (.50) | / |
| B | 0 (.0) | 2 (.06) | 12 (.10) | / |
| C | 0 (.0) | 0 (.0) | 8 (.7) | / |
| Unreported | 60 (.100) | 15 (.48) | 40 (.33) | / |
| Pathology** | 42 (.70) | / | / | / |
| Well differentiated HCC | 2 (.03) | / | 44 (.36.36) | / |
| Well to moderately differentiated HCC | 5 (.08) | / | / | / |
| Moderately differentiated HCC | 19 (.32) | / | 18 (14.88) | / |
| Moderately to poorly differentiated HCC | 12 (.20) | / | / | / |
| Poorly differentiated HCC | 4 (.07) | / | 4 (3.31) | / |
| Unknown differentiation HCC | 0 (.0) | / | 55 (45.45) | / |
| Non-HCC lesions | 4 (.07) | / | / | / |
| -Malignant | 2 (.03) | 8 (.1) | / | / |
| Cholangiocarcinoma*** | 1 (.02) | 7 (22.58) | / | / |
| Metastasis | 1 (.02) | 1(3.23) | / | 291 (1.00) |
| -Pre-malignant | 0 (.0) | 20 (64.52) | / | / |
| Dysplastic nodule | 0 (.0) | 14 (45.16) | / | / |
| Regenerative nodule | 0 (.0) | 6 (19.35) | / | / |
| -Benign | 2 (.03) | 3 (9.68) | / | / |
| Hemangioma | 0 (.0) | 2 (6.45) | / | / |
| FNH | 2 (.03) | 1 (3.2) | / | / |
| Unreported | 14 (.23) | 0 | / | / |
Note.—
data are expressed in percentages. Etiology of cirrhosis are non-mutually exclusive. NASH = nonalcoholic steatohepatitis, HIV = Human Immunodeficiency Virus.
Data are number of patients; data in parentheses are percentages for each group (HCC and Non−HCC).
including one case of hepatocholangio-carcinoma.
FNH = focal nodular hyperplasia, /: not applicable.
2.2. Reference standard for contrast-enhancement phases
The reference standard and output to be predicted was a time interval spanned by five clinically utilized contrast-enhancement phases. The acquisition times of the dynamic CT-scan images were continuous variables that we transformed into a discrete categorization over 5 phases: (i) Non-contrast [NCP]; (ii) Early Arterial [E-AP]; (iii) Optimal Arterial [O-AP]; (iv) Optimal Portal Venous [O-PVP]; (v) Late Portal Venous [L-PVP]. At each time-point the original time values (i.e., the time after contrast injection provided in the image Digital Imaging and Communications in Medicine DICOM header) were replaced by a single discrete value (i.e., 1–5) representative of a time interval spanned by the 5 phases. Appendix, II (Fig. 2).
Fig. 2.
Flow diagram.
2.3. Input of the phase classifier
The input of the phase classifier was the mean density of the abdominal aorta and the portal vein measured from a single abdominal Computed Tomography image in each patient at each timepoint. Radiologists who were blinded to the assessment of the predicted outcome drew regions of interest (ROIs) in the abdominal aorta and portal vein, upon which the mean densities were computed. 2D ROI of approximately 2 cm of longest diameter were manually segmented in the lumen of aorta and portal vein. The abdominal aorta and the portal vein were selected because they were shown to be the two most informative locations for predicting the contrast-enhancement phase [2] (Appendix, III).
2.4. CECT-QC algorithm: development and validation in cohort A
The CECT-QC algorithm was trained in the training-set (48 patients, 1884 time-points) and tested in the test-set (12 patients, 475 time-points). Each patient had on average 39 consecutive acquisition time-points from the time of contrast-enhancement injection to the portal venous phase.
The CECT-QC algorithm was developed in the training-set. A random forest classifier was trained to classify each conventional CT image acquired at a single time point into one of the 5 phases using as input the density in the aorta and portal vein on this single image at a single time point. During the training, each time-point from one patient was treated as an independent time-point. The random forest classifier utilized ten trees with a maximum growing depth of five (22). The output of the classifier was one of the 5 phases.
The accuracy of the developed CECT-QC algorithm was evaluated against the test-set. The performance was calculated using 95 % confidence intervals [95 % CIs] using 1000-bootstrap [22]. A normalized confusion matrix displays the comparison of the predicted phase using the classifier to the reference standard phase.
2.5. CECT-QC algorithm: clinical applications
The CECT-QC algorithm was evaluated in three independent cohorts (Appendix, I) to assess whether it could affect clinical practice. Blinded to clinical data/outcomes, the participating radiologists delineated the two predefined anatomical landmarks, an ROI in aorta and an ROI in portal vein, to compute average densities of aorta and portal vein. When studying effects of phase timing on tumor imaging biomarkers, we segmented each liver lesion using a previously described platform [23].
2.6. CECT-QC algorithm: clinical applications in cirrhotic patients (cohorts B and C)
The CECT-QC algorithm was applied to cohorts B and C totaling 152 cirrhotic patients with triphasic CT-scans (456 time points), collected from 27 different institutions (Appendix, I). First, we evaluated the percentages of patients showing O-AP and O-PVP. Second, we evaluated liver nodule densities at NCP, E-AP, O-AP, O-PVP, and L-PVP. Third, the ‘washout’ as measured by the change in tumor density between the arterial phase and the portal venous phase was compared between (a) patients with both O-AP and O-PVP and (b) patients with at least one non-optimal acquisition. Fourth, we determined whether there was an association between the phase predicted by the CECT-QC algorithm and the ‘washout’ radiomic phenotype.
‘Washout’ radiomic features characterized the change in radiomic features between the AP and the PVP (delta features). The kinetics of the density of HCC liver nodules indeed has a washout pattern that is distinct from non−HCC liver nodules [24]. To this end, we extracted 1160 radiomic features from each liver nodule at all 456 time points (Appendix, III). Radiomics features extraction used a previously published methodology [8]. Using unsupervised agglomerative clustering [25], ‘washout’ radiomic phenotypes were divided into clusters. We determined whether clusters were related to histology (HCC vs. non−HCC) or acquisition of technical parameters (CECT phase) (Appendix, III).
2.7. CECT-QC algorithm: clinical applications in non-cirrhotic patients (cohort D)
First, we evaluated the percentage of patients with O-AP and O-PVP and compared the results to a previously described method [2]. Second, we evaluated liver nodule densities at NCP, E-AP, O-AP, O-PVP, and L-PVP (Appendix, IV).
2.8. Statistical analysis methods
Statistical analyses were conducted using Python3.6, Matlab2016a and SPSS24.0. The reported statistical significance levels were two-sided, with statistical significance set at P = 0.05 using the Bonferroni correction. The accuracy of the CECT-QC algorithm was evaluated in the test-set which consisted of unseen data by computing the percentage of patients correctly classified as compared to the reference standard. ANOVA compared the distribution of liver lesion densities in different groups and in different CE phases.
3. Results
3.1. Patient cohorts
Characteristics of the patients and the technical settings of CT scans in cohorts A, B, C, and D are provided in Tables 1 and E1.
3.2. Reference standard
Figs. 3A and E1 show the aggregated densities in the aorta and portal vein as a function of a continuous time after injection. Using this information, a discrete classification into 5 phases was performed for all patients. There was a significant inter-patient variability in contrast-enhancement biodistribution kinetics (Appendix, II and Table E2).
Fig. 3.
Training set: Reference standard (A) and model building (B). (A) Density distributions of aorta and portal vein as a function of normalized time to peak (NTTP, Appendix, II) in cohort A. The solid lines are mean values and the dashed lines are 95 % CIs. The five phases were defined according to the criteria in Table 2. (B) The model uses the density in the aorta and portal vein to predict the phase as defined by an NTTP: (i) Black dots: Non contrast phase; (ii) Blue dots: Early arterial phase; (iii) Green dots: Optimal arterial phase; (iiii) Yellow dots: Optimal portal venous phase; (v) Grey dots: Late portal venous phase. The conventional CT only contains one time-point image, so we built our model based on image densities of aorta and portal vein.
3.3. CECT-QC algorithm: Development and validation in cohort A
Performance of the CECT-QC algorithm (Random Forest classifier) was evaluated in the test-set. The CECT-QC algorithm reached an overall accuracy of 79.4 % [95 % CI = 75.2 %, 82.9 %] (377 correct predictions out of 475 total predictions) to predict the phase based on the analysis of a single image at a single time point. Fig. 4 shows the normalized confusion matrix of prediction: the accuracies to detect optimal acquisition phases were 98 % for NCP, 90 % for O-AP, and 84 % for O-PVP.
Fig. 4. This confusion matrix compares the true phase and predicted phase in the test-set.
The CE phase can be predicted by machine-learning with good accuracy. The only difficulty was identification of late PVP which can be attributed to the fact that contrast enhancement reaches a plateau phase (Fig. 3A).
3.4. CECT-QC algorithm: clinical applications in cirrhotic patients (cohorts B and C)
The percentage of cases with optimal CE phases was evaluated. CT-scans labelled as Non-Contrast, AP or PVP were respectively classified as NCP in 98 % (n = 149/152), O-AP in 95.4 % (n = 145/152), and O-PVP in 41.4 % (n = 63/152) of patients (Tables 2 and E3).
Table 2. Distribution of CE phases in cohorts B, C and D.
98.0 % of non-contrast enhanced CTs are classified by the algorithm as optimal NC. 95.4 % of arterial phase enhanced CTs are classified by the algorithm as optimal AP. Only 41.4 % of portal venous phase enhanced CTs are classified by the algorithm as optimal PVP.
| Cohorts B and C |
Cohort D | |||
|---|---|---|---|---|
| Intended acquisition phase | Non-contrast | Arterial | PV | PV |
| Predicted phase | ||||
| Non-contrast | 149 (98.0)* | 0 (0) | 1 (0.7) | 0 (0) |
| Early AP | 0 (0) | 0 (0) | 0 (0) | 0 (0) |
| Optimal AP | 1(0.7) | 145 (95.4)* | 5 (3.3) | 102 (17.5) |
| Optimal PVP | 0 (0) | 2 (1.4) | 63 (41.4)* | 301 (51.7)* |
| Late PVP | 2(1.3) | 5 (3.2) | 83 (54.6) | 179 (30.7) |
Note. Data are number of patients; data in parentheses are percentages. PV phase: portal venous phase.
Optimal acquisition. In cohort D, 291 patients had two contrast-enhanced portal venous phase CT-scan acquisitions per patient (582 CTs: baseline and first follow-up).
Liver nodule density was significantly different between the phases predicted by the CECT-QC algorithm (P < 10−70). The liver nodule density (mean ± SD, in HU) peaked at the optimal-PVP (91 ± 20 HU), showing 47 % and 15 % increases in tumor density as compared to optimal-AP (62 ± 18 HU) and late-PVP (71 ± 21 HU) respectively (Table 3).
Table 3.
Tumor density according to each predicted phase according to the CECT-QC algorithm in cohorts B, C and D.
| Predicted phase | Abbreviation | Tumor Density (average ± SD) |
|
|---|---|---|---|
| Cohort B and C (P-value = 2.6*10−71) | Cohort D (P-value = 1.8*10−18) | ||
| Non-contrast | NCP | 39 ± 11 | None |
| Early AP | E-AP | None | None |
| Optimal AP | O-AP | 62 ± 18 | 61 ± 14 |
| Optimal PVP | O-PVP | 91 ± 20 | 73 ± 15 |
| Late PVP | L-PVP | 71 ± 21 | 62 ± 14 |
Note.—SD: standard deviation.
The ‘washout’ in tumor density between AP and PVP was significantly higher in patients with both optimal-AP and optimal-PVP (mean washout: +26.5 HU, n = 59 patients), than patients with at least one non-optimal acquisition (mean washout: +12.0 HU, n = 93 patients) (P = 10−8). Overall, the area under the curve (AUC) of ‘washout’ in tumor density for the diagnosis of HCC was 0.61 (95 % CI: [0.50 – 0.71]) (Figs. 5 and 6, E3).
Fig. 5. Twelve percent of washout features were sensitive to contrast-enhancement quality.
Heatmap of 1160 delta Radiomic features (features in PVP minus features in AP). Each feature was normalized by a z-score. The patients were divided into five groups by unsupervised Agglomerative clustering. For each patient, delta aorta density, delta PV density, timing at arterial phase CT, timing at portal venous CT, both optimal timing, and pathology are shown below with corresponding color codes. The radiomic feature clusters are significantly correlated with contrast-enhancement quality (optimal AP and optimal PVP vs. non-optimal AP or non-optimal PVP, P = 0.0003) but are not correlated with nodule biology (HCC vs. non−HCC, P = 0.12).
Fig. 6. Comparison of the biological signal to the contrast-enhancement noise using the CECT-QC algorithm.
Fig. 6 For 1160 features, the ratio between the biological signal (difference between HCC and non HCC) and the contrast-enhancement noise (difference between optimal and non optimal acquisitions) was computed. The value is displayed in A. The waterfall plots (B, C, D) display the ratio for the 1160 features at the AP (B), the PVP (C) and the washout: difference between AP and PVP (C).
Unsupervised clustering of liver nodule phenotypes using 1160 radiomic features in 152 cirrhotic patients (non−HCCs: 31, HCC: 121) identified five clusters (Figs. 5 and 6, E3). These clusters appeared to be more related to contrast-enhancement quality (liver nodule phenotypes at optimal CE phase vs. non-optimal CE phase, P = 0.0003) than nodule biology (HCC phenotype vs. non−HCC phenotype, P = 0.12). This is further demonstrated by the fact that 12 % (n=140/1160) of the studied radiomic features were significantly different when they were extracted/computed from the abdominal CT images acquired at optimal and non-optimal phases (ANOVA, P: 0.048 to 10−7).
3.5. CECT-QC algorithm: clinical applications in non-cirrhotic patients (cohort D)
The percentage of optimal CE phases was evaluated. PVP acquisitions were optimal in 51.7 % of examinations (n = 301/582) (Table 2). Non-optimal PVP phases were associated with a 15 % decrease in tumor density (Table 3).
Using a confusion matrix (Fig. E2), we compared the agreement between the CECT-QC algorithm (5 phases classifier) and a previously published method (3 phases classifier without NCP and early-AP) [2]. The agreement was 92 % for optimal-AP, 74 % for optimal-PVP and 81 % for late-PVP. Of note, cohort D was intentionally collected as the portal venous phase, hence the absence of occurrence of NCP and Early-AP.
The liver metastasis density (mean ± SD in HU) peaked at optimal-PVP (61 ± 14 HU). Moreover, optimal PVP showed increases of 20 % and 18 % in tumor density compared to optimal-AP and late-PVP respectively (Table 3).
3.6. Code sharing
An excel spreadsheet will be provided upon request to calculate the CECT phase. Once the input is provided (the density in the aorta and in the portal vein), the excel spreasheet automatically generate the CECT phase. The output of this machine-learning algorithm can therefore be used for radiomics-based precision medicine.
4. Discussion
Using dynamic CT-scan images and a machine-learning method, we developed a CECT-QC algorithm to identify the contrast-enhancement phase based on analysis of the densities of the aorta and of the portal vein on a single, standard-of-care, CT-scan image. This tool allows recognition of a CT-scan image acquired at Non-contrast, Early-Arterial, Optimal-Arterial, Optimal-PV or Late-PV phases.
The CECT-QC algorithm was validated on data collected over a decade in multiple institutions, generalizable across vendors, and applicable to all contrast-enhancement phases. Because the CECT-QC algorithm is based on routinely acquired CT scans, and on the density measurements in the aorta and portal vein, it can be widely incorporated into clinical practice or a radiomic precision medicine program without any additional costs.
Results show that the CECT-QC algorithm can be useful to appraise the quality of the contrast-enhancement acquisition since non-optimal acquisition phases are frequent, particularly during the PVP. We observed that most arterial phases are optimal, while optimal PVPs were achieved in only 41 % of cirrhotic patients and 52 % of non-cirrhotic colorectal cancer patients with liver metastases. These results in cirrhotic patients may be explained by several factors. Cirrhosis alters the biodistribution of contrast-enhancement product. Hepatic architecture distortion with sinusoidal microcirculatory dysfunction leads to portal hypertension, ultimately leading to liver failure [26–28]. Cirrhotic cardiomyopathy alters the cardiac output and contractility, modifying contrast-enhancement kinetics [29]. Finally, portal hypertension-related abnormal hemodynamic distribution and cirrhotic cardiomyopathy lead to marked changes in systemic vascular resistance, expanded blood volume and splanchnic bed widening [29].
The clinical utility of this CECT-QC algorithm for radiomic-based precision diagnosis and treatment was validated in the most common primary and metastatic liver malignancies (HCC and colorectal cancer) in patients with normal and impaired liver function (cirrhotic patients).
This study demonstrated that the CECT-QC algorithm is useful for radiomic-based precision diagnosis. Arterial hyperenhancement and washout are used to diagnose HCC. However, this study shows the considerable variability in tumor density caused by variability in contrast-enhancement quality. Our results suggest that indeterminate liver nodule diagnoses may be altered by poor quality in contrast-enhancement. A sizable number (140 of 1160, 12 %) of the quantitative radiomic features deciphering the phenotype of indeterminate cirrhotic liver nodules was influenced by the quality of the contrast-enhancement phase. Unsupervised clustering showed that imaging phenotype of indeterminate liver nodules in cirrhotic patients was related to CE quality (optimal vs. non-optimal examinations) rather than histology (HCC vs. Non−HCC). This point highlights that the very distinctive phenotype of HCC may not be observed in some indeterminate liver nodules because of inadequate CE phase.
This study showed that the CECT-QC algorithm is useful for radiomic-based precision treatment. Contrast-enhancement phase is indeed a confounding variable when assessing tumor biology using density as a biomarker. We identified clinically significant differences in liver lesion density between optimal and non-optimal acquisitions. This is clinically relevant since tumor density is a surrogate biomarker of tumor vascularity, with wide-ranging applications across cancer diagnoses and therapies. Density is a radiological biomarker used for response assessment in both clinical routine and research for both primary and secondary liver malignancies (CHOI and MASS criteria).
In the emerging field of radiomics, extraction of the tumors’ quantitative imaging biomarkers must take into consideration the feature’s robustness regarding contrast-enhancement, or minimize the contrast-enhancement variability. Our CECT-QC algorithm could be incorporated into all radiomic pipelines studying abdominal CECT for (i) selecting only optimally acquired images, and (ii) correcting or normalizing the inherent phase-induced variability.
The main limitation of this study is the overlap in densities between the five, defined phases. The acquisition time of an image is a continuous variable that was transformed into a single discrete value (phase) representative of a time interval. The separation of overlapping time points into five distinct timing categories was a technical challenge and left to the discretion and experience of the primary investigators, as there is no consensus for the definition of the five phases’ boundaries (i.e., start and end times). In order to define the boundaries for these phases, data from the existing literature as well as aggregated data from a significant number of patients and time points (cohort A, training-set) were collected. Moreover, these phases are used in clinical routine in Radiology, showing clinical significance. Additionally, the CECT-QC algorithm showed a lower accuracy for distinguishing late-PVP from O-PVP. This is because the densities in the aorta and in the portal vein reach a plateau at the end of O-PVP and at the beginning of late-PVP, resulting in an overlap between these phases. There are opportunities to improve the CECT-QC algorithm in order to better distinguish these two phases by adding ROIs in other organs, such as the kidney, to appraise the elimination of the contrast-enhancement product at late-PVP. This was out of the scope of this manuscript because the kidney was not included in the field of view in many of the available dynamic CT-scan studies. Liver parenchyma density information was not used to build the classifier since previous studies demonstrated that it does not provide incremental value [2] and that the estimates of liver density can be decreased by steatosis and chemotherapy.
5. Conclusion
In conclusion, the CECT-QC algorithm developed in this study was trained and validated in a large dataset of 503 patients and 3397 time-points. It offers a proof of concept that machine-learning, applied to a single image using two simple anatomical landmarks, allows an accurate categorization of five contrast-enhancement phases. Using this quality control tool can help to improve the reproducibility of tumor imaging biomarker extraction for radiomic-based precision diagnosis and treatment [30,31,32] in liver diseases, both at patient and clinical trial level.
Supplementary Material
Funding
National Cancer Institute (NCI) supported this work in part by Grant U01 CA225431. The content is solely the responsibility of the authors and does not necessarily represent the views of the funding source.
The work was partially funded by a grant from Fondation Philanthropia (LD), Fondation Nuovo-Soldati (LD), and through research grants from Alain Rahmouni French Society of Radiology-CERF 2018 (FZM).
Abbreviations:
- AP
arterial phase
- CE
contrast enhanced
- CT
Computed Tomography scan
- HCC
Hepatocellular carcinoma
- NCP
non-contrast phase
- PVP
portal venous phase
- QC
quality control
Footnotes
Guarantor
The scientific guarantor of this publication is Laurent Dercle.
Informed consent
Written informed consent was waived by the Institutional Review Board.
Ethical approval Institutional Review Board approval was obtained from participating institutions.
Methodology
• retrospective.
• diagnostic.
• multicenter study.
Declaration of Competing Interest
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.
Appendix A. Supplementary data
Supplementary material related to this article can be found, in the online version, at doi:https://doi.org/10.1016/j.ejrad.2020.108850.
References
- [1].Lambin P, Leijenaar RT, Deist TM, et al. , Radiomics: the bridge between medical imaging and personalized medicine, Nat. Rev. Clin. Oncol 14 (749) (2017). [DOI] [PubMed] [Google Scholar]
- [2].Dercle L, Lu L, Lichtenstein P, et al. , Impact of variability in portal venous phase acquisition timing in tumor density measurement and treatment response assessment: metastatic colorectal cancer as a paradigm, JCO Clin. Cancer Inform 1 (2017) 1–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [3].European Association for the Study of the Liver. Electronic address eee, European Association for the Study of the L, EASL clinical practice guidelines: management of hepatocellular carcinoma, J. Hepatol (2018), 10.1016/j.jhep.2018.03.019. [DOI] [PubMed] [Google Scholar]
- [4].El-Serag HB, Rudolph KL, Hepatocellular carcinoma: epidemiology and molecular carcinogenesis, Gastroenterology 132 (2007) 2557–2576. [DOI] [PubMed] [Google Scholar]
- [5].Dercle L, Billey C, Cognet T, et al. , Spleno-hepatic index to predict portal hypertension by equilibrium radionuclide ventriculography, Nucl. Med. Commun 39 (2018) 1138–1142. [DOI] [PubMed] [Google Scholar]
- [6].Kim MY, Jeong WK, Baik SK, Invasive and non-invasive diagnosis of cirrhosis and portal hypertension, World J. Gastroenterol 20 (2014) 4300. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [7].Talakić E, Schaffellner S, Kniepeiss D, et al. , CT perfusion imaging of the liver and the spleen in patients with cirrhosis: Is there a correlation between perfusion and portal venous hypertension? Eur. Radiol 27 (2017) 4173–4180. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [8].Mokrane FZ, Lu L, Vavasseur A, et al. , Radiomics machine-learning signature for diagnosis of hepatocellular carcinoma in cirrhotic patients with indeterminate liver nodules, Eur. Radiol (2019), 10.1007/s00330-019-06347-w. [DOI] [PubMed] [Google Scholar]
- [9].Llovet JM, Bru C, Bruix J, Prognosis of hepatocellular carcinoma: the BCLC staging classification, Semin. Liver Dis 19 (1999) 329–338. [DOI] [PubMed] [Google Scholar]
- [10].Di Martino M, De Filippis G, De Santis A, et al. , Hepatocellular carcinoma in cirrhotic patients: prospective comparison of US, CT and MR imaging, Eur. Radiol 23 (2013) 887–896. [DOI] [PubMed] [Google Scholar]
- [11].Ronot M, Fouque O, Esvan M, Lebigot J, Aube C, Vilgrain V, Comparison of the accuracy of AASLD and LI-RADS criteria for the non-invasive diagnosis of HCC smaller than 3cm, J. Hepatol (2017), 10.1016/j.jhep.2017.12.014. [DOI] [PubMed] [Google Scholar]
- [12].Serste T, Barrau V, Ozenne V, et al. , Accuracy and disagreement of computed tomography and magnetic resonance imaging for the diagnosis of small hepatocellular carcinoma and dysplastic nodules: role of biopsy, Hepatology 55 (2012) 800–806. [DOI] [PubMed] [Google Scholar]
- [13].Forner A, Vilana R, Ayuso C, et al. , Diagnosis of hepatic nodules 20 mm or smaller in cirrhosis: prospective validation of the noninvasive diagnostic criteria for hepatocellular carcinoma, Hepatology 47 (2008) 97–104. [DOI] [PubMed] [Google Scholar]
- [14].Freeman Richard B, Mithoefer A, Ruthazer R, et al. , Optimizing staging for hepatocellular carcinoma before liver transplantation: a retrospective analysis of the UNOS/OPTN database, Liver Transplant 12 (2006) 1504–1511. [DOI] [PubMed] [Google Scholar]
- [15].Edeline J, Boucher E, Rolland Y, et al. , Comparison of tumor response by Response Evaluation Criteria in Solid Tumors (RECIST) and modified RECIST in patients treated with sorafenib for hepatocellular carcinoma, Cancer 118 (2012) 147–156 %@ 0008–0543X. [DOI] [PubMed] [Google Scholar]
- [16].Llovet JM, Ricci S, Mazzaferro V, et al. , Sorafenib in advanced hepatocellular carcinoma, N. Engl. J. Med 359 (2008) 378–390 %@ 0028–4793. [DOI] [PubMed] [Google Scholar]
- [17].Brufau BP, Cerqueda CS, Villalba LB, Izquierdo RS, Gonzalez BM, Molina CN, Metastatic renal cell carcinoma: radiologic findings and assessment of response to targeted antiangiogenic therapy by using multidetector CT, Radiographics 33 (2013) 1691–1716. [DOI] [PubMed] [Google Scholar]
- [18].Benjamin RS, Choi H, Macapinlac HA, et al. , We should desist using RECIST, at least in GIST, J. Clin. Oncol 25 (2007) 1760–1764. [DOI] [PubMed] [Google Scholar]
- [19].Choi H, Charnsangavej C, Faria SC, et al. , Correlation of computed tomography and positron emission tomography in patients with metastatic gastrointestinal stromal tumor treated at a single institution with imatinib mesylate: proposal of new computed tomography response criteria, J. Clin. Oncol 25 (2007) 1753–1759. [DOI] [PubMed] [Google Scholar]
- [20].Choi H, Charnsangavej C, de Castro Faria S, et al. , CT evaluation of the response of gastrointestinal stromal tumors after imatinib mesylate treatment: a quantitative analysis correlated with FDG PET findings, AJR Am. J. Roentgenol 183 (2004) 1619–1628. [DOI] [PubMed] [Google Scholar]
- [21].Lencioni R, Llovet JM, Modified RECIST (mRECIST) Assessment for Hepatocellular Carcinomaseminars in Liver Disease, © Thieme Medical Publishers, 2010, pp. 052–060. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [22].Stehman SV, Selecting and interpreting measures of thematic classification accuracy, Remote Sens. Environ 62 (1997) 77–89. [Google Scholar]
- [23].Yang H, Schwartz LH, Zhao B, A response assessment platform for development and validation of imaging biomarkers in oncology, Tomography 2 (2016) 406–410. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [24].Forner A, Vilana R, Ayuso C, et al. , Diagnosis of hepatic nodules 20 mm or smaller in cirrhosis: prospective validation of the noninvasive diagnostic criteria for hepatocellular carcinoma, Hepatology 47 (2008) 97–104. [DOI] [PubMed] [Google Scholar]
- [25].Rokach L, Maimon O, Clustering methodsData Mining and Knowledge Discovery Handbook, Springer, 2005, pp. 321–352. [Google Scholar]
- [26].Bosch J, Berzigotti A, Garcia-Pagan JC, Abraldes JG, The management of portal hypertension: rational basis, available treatments and future options, J. Hepatol 48 (Suppl 1) (2008) S68–92. [DOI] [PubMed] [Google Scholar]
- [27].Sanyal AJ, Bosch J, Blei A, Arroyo V, Portal hypertension and its complications, Gastroenterology 134 (2008) 1715–1728. [DOI] [PubMed] [Google Scholar]
- [28].Tsochatzis EA, Bosch J, Burroughs AK, Liver cirrhosis, Lancet 383 (2014) 1749–1761. [DOI] [PubMed] [Google Scholar]
- [29].Zardi EM, Abbate A, Zardi DM, et al. , Cirrhotic cardiomyopathy, J. Am. Coll. Cardiol 56 (2010) 539–549. [DOI] [PubMed] [Google Scholar]
- [30].Dercle L, Lu L, Schwartz LH, Qian M, Tejpar S, Eggleton P, Zhao B, Piessevaux HJ, Radiomics response signature for identification of metastatic colorectal cancer sensitive to therapies targeting egfr pathway, Natl. Cancer Inst (February) (2020), 10.1093/jnci/djaa017 pii: djaa017. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [31].Rozenblum L, Mokrane FZ, Yeh R, Sinigaglia M, Besson FL, Seban RD, Zadro C, Dierickx L, Chougnet CN, Partouche E, Revel-Mouroz P, Zhao B, Otal P, Schwartz LH, Dercle L, Imaging-guided precision medicine in non-resectable gastro-entero-pancreatic neuroendocrine tumors: a step-by-step approach, Eur. J. Radiol 122 (January) (2020) 108743, 10.1016/j.ejrad.2019.108743 Epub 2019 Nov 12. Review. [DOI] [PubMed] [Google Scholar]
- [32].Tselikas L, Sun R, Ammari S, Dercle L, Yevich S, Hollebecque A, Ngo-Camus M, Nicotra C, Deutsch E, Deschamps F, de Baere T, Role of image-guided biopsy and radiomics in the age of precision medicine, Chin. Clin. Oncol 8 (Dec. (6)) (2019), 10.21037/cco.2019.12.02. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.