Machine learning-based Radiomics analysis of preoperative functional liver reserve with MRI and CT image

doi:10.21203/rs.3.rs-2309913/v1

Download PDF

Research Article

Machine learning-based Radiomics analysis of preoperative functional liver reserve with MRI and CT image

https://doi.org/10.21203/rs.3.rs-2309913/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 17 Jul, 2023

Read the published version in BMC Medical Imaging →

You are reading this latest preprint version

Objective: Comparing indocyanine green retention rate at 15 min (ICG-R15) can accurately evaluate functional liver reserve, we investigated the ability of Gd-EOB-DTPA-enhanced hepatic magnetic resonance imaging (MRI) and contrast-enhanced computed tomography (CT) image in hepatocellular carcinoma (HCC) patients’ radiomics models for evaluation of functional liver reserve. To assist doctors in evaluating hepatic functional reserve in the hospitalthat lacks expensive ICG equipment.

Methods: 190 HCC patients in total were retrospectively enrolled and randomly classified into a training dataset (CT: n = 152, MR: n = 90) and a test dataset (CT: n = 38, MR: n =22). Then, radiomics features from MRI and CT images were extracted. The features associated with the ICG-R15 classificationwere picked out. Six machine learning (ML) classifiers were used for the ML-model investigation, and the accuracy (ACC) and area under ROI curve (AUC) of receiver operating characteristic (ROC) with 95% confidence intervals (CI) utilized for ML-model performance evaluation.

Results: 107 different radiomics features were extracted from MRI and CT respectively. The features related to ICG-R15 classification were selected. In MRI groups, when ICG-R15=10% was selected as a threshold, classifier LightGBM performed best for its AUC was 0.932 and ACC 0.955. When ICG-R15=20%, classifier LightGBM performed best for its AUC was 0.938 and ACC 0.913. When ICG-R15=30%, classifier XGBoost performed best for its AUC was 0.972 and ACC 0.955. For CT groups, when ICG-R15=10% was selected as a threshold, classifier LightGBM performed best for its AUC was 0.891 and ACC 0.868. When ICG-R15=20%, classifier SVM performed best for its AUC was 0.877 and ACC 0.842. When ICG-R15=30%, classifier LightGBM performed best for its AUC was 0.927 and ACC 0.947.

Conclusions:Both the MRI and CT machine learning models are considered valuable noninvasive methods for the evaluation of functional liver reserve. The performance of the MRI model was better than that of the CT model in the assessment of functional liver reserve.

Radiomics

Functional liver reserve

Machine learning

Gd-EOB-DTPA-enhanced hepatic MRI

contrast-enhanced CT

Primary liver cancer is the sixth most commonly diagnosed cancer and the third contributing factor for global cancer death, with about 906,000 new cases and 830,000 deaths in 2020 [1]. Hepatocellular carcinoma (HCC) comprises 75%-85% of primary liver cancer [1]. Asia and Africa get the highest incidence of HCC in the world [2]. Most patients are diagnosed in advanced stages because of insidious symptoms [3], while partial hepatectomy (PH) is still the optimal HCC therapy [4]. However, one important complication is the post-hepatectomy liver failure (PHLF), which is the major cause of postoperative mortality. As a result, preoperative evaluation of functional liver reserve is necessary for HCC patients because most of the patients are complicated with hepatitis or liver cirrhosis.

The incidence of PHLF is 0.7%-9.1%, while it reaches as high as 58.22% in major hepatectomy [5–6]. Presurgical evaluation of functional liver reserve is critical to evaluate the risk of hepatectomy and avoid PHLF. In clinical work, liver volumetry and scoring systems based on blood tests, such as ALB, AST, TBIL, and indocyanine green retention at 15 minutes retention rate (ICG-R15) are classic indexes for the evaluation of functional liver reserve. The 3D reconstruction technology has become a conventional method for liver volumetry [7]. MELD score, Child-Turcotte-Pugh (CTP) score, and Albumin-bilirubin (ALBI) grade are well-utilized scoring systems. By comparing these indexes, the ICG-R15 can help doctors detect functional liver reserve abnormality earlier and more accurately. ICG-R15 has also been proven to have a positive correlation to liver failure and morbidity after hepatectomy [8] Thus, the ICG clearance test was considered an evaluation of preoperative function liver reserve [9].

Radiomics, which is one emerging methodology in medicine, depicts quantitative computerized algorithm-based features extracted from traditional image materials like CT or MRI images [10–12]. The relation of medical images to clinical manifestations can be identified via machine learning (ML) approaches [13, 14]. In the past study, the applications of radiomics in liver diseases were reported in several aspects of investigating diagnosis, staging, liver tumor biological behaviors, and prognosis of primary liver cancer [15, 16]. Nevertheless, the value of radiomics in evaluating functional liver reserve has been scarcely examined. The comparison of radiomic model based on Gadolinium ethoxybenzyl dimeglumine (Gd-EOB-DTPA)-enhanced hepatic MRI (MRI) and contrast-enhanced CT (CT) has not been also reported.

Under this direction, in this work, it was evaluated whether it is possible to carry out an assessment of functional liver reserve by using radiomics models derived from Gd-EOB-DTPA-enhanced hepatic MRI or contrast-enhanced CT compared with ICG-R15 in HCC patients. The validity between MRI and CT images was also compared.

Flowchart

According to the flowchart of this work (Fig. 1), the Gd-EOB-DTPA-enhanced hepatic MRI (MRI) data from 112 patients and contrast-enhanced CT (CT) data from 190 patients were retrospectively collected. The hepatobiliary phase at 15 minutes of MRI and the portal venous phase of CT were selected, the ROI liver region was segmented, and the features were extracted from the image. Next, the features were further screened through P-value and correlation coefficient. The dataset was randomly divided into a training dataset and a test dataset, the training dataset aimed to train the model, while the features associated with the label of ICG-R15 were screened by using the least absolute shrinkage and selection operator (LASSO) algorithm. Multiple machine learning algorithms were also used for training on the training dataset and predicting functional liver reserve classification results on the test dataset. At last, the best model was selected to evaluate functional liver reserve classification.

Patients

By reviewing the patients who were diagnosed with HCC in our hospital from May 2017 to April 2022, he inclusion criteria were the following: (1) all patients were confirmed to have HCC; (2) Gd-EOB-DTPA-enhanced hepatic MRI or contrast-enhanced CT in all phases was completed one week before treatment or surgery; (3) ICG clearance test within one week before treatment or surgery; (4) patients without jaundice during ICG clearance test [17]; (5) all patients had no history of previous liver surgery. A total of 190 patients were included in this study. Details are listed in Table 1. The Ethics Committee of the Affiliated Hospital of Qingdao University approved this study with the ethical approval number QYFY-WZLL-27465.

Table 1

Demographics and preoperative data of patients
Characteristic	Categories	Value		Number
Characteristic	Categories	MRI	CT	MRI	CT
dataset	Training dataset (80%)			89	152
	Test dataset (20%)			23	38
Age (mean ± SD)		58.62 ± 8.62 y	59.42 ± 8.93 y
Gender	Female			25	44
	Male			87	146
HBV infection	Yes			104	168
	No			8	22
Liver cirrhosis	Yes			75	122
	No			37	68
BMI (mean ± SE)		25.02 ± 0.39 kg/m2	24.75 ± 0.27 kg/m2
TBIL (mean ± SE)		21.32 ± 1.60 µmol/L	22.37 ± 1.10 µmol/L
ALB (mean ± SE)		36.26 ± 0.66 g/L	36.99 ± 0.43 g/L
ALT (mean ± SE)		48.38 ± 10.21 IU/L	49.06 ± 5.96 IU/L
PT (mean ± SE)		11.34 ± 0.14s	11.45 ± 0.11s
AST (mean ± SE)		52.31 ± 6.81 IU/L	56.08 ± 5.24 IU/L
GGT (mean ± SE)		99.44 ± 16.15 IU/L	108.12 ± 12.23 IU/L
ICG-R15	ICG-R15 ≤ 10% vs ICG-R15༞10%			62vs128	45vs67
	ICG-R15 ≤ 20% vs ICG-R15༞20%			126vs64	78vs34
	ICG-R15 ≤ 30% vs ICG-R15༞30%			160vs30	93vs19
ALT: alanine transaminase, BMI: Body Mass Index, HBV: Hepatitis B Virus; TBIL: total Bilirubin, ALB: albumin, PT: prothrombin time, AST: aspartate aminotransferase, GGT: gamma-glutamyltransferase, ICG-R15: indocyanine green retention rate at 15 min, y: years.

ICG Clearance Test

After 6 hours of fasting, the patient was in a supine position and injected 0.5 mg/kg ICG (Dandong Yichuang Pharmaceutical Co., Ltd., Liaoning, China) intravenously into a peripheral vein within 10 seconds. The ICG retention rate was measured with ICG pulse spectrophotometry (DDG 3300K, Japan) after 15 minutes of injection. The ICG-R15 value was expressed as the percentage of ICG retention in serum 15 minutes after injection.

MRI and CT acquisition

The MRI examination was conducted by using a Siemens Verio-Dot 3.0 T MRI scanner. The CT examination was conducted by using a Siemens SOMATOM Definition Flash scanner. From the top of the liver to the pelvis, scans were performed. The MRI scanning parameters were selected as below: the scanning layer thickness was 3 mm, the interval was 1 mm, the matrix was 182 × 320, and the field of view was 40 cm × 40 cm. After the elbow vein injection of 20 mL Gd-EOB-DTPA (with flow rate of 2.0 ml/s and dose of 0.025 mmol/kg), enhanced scans was conducted on the patients. The delay times for the portal and hepatobiliary phases were 7 s and 20 min respectively. The CT scanning parameters were selected as below: voltage of 120 kV, current of 200–350 mA, scanning layer thickness of 1 mm, layer spacing of 5 mm, and matrix of 512 ×512. After the peripheral vein injection of iohexol containing 350 mg/ml of iodine (with flow rate of 3.0 ml/s and dose of 1.5 ml/kg), enhanced scans was conducted on the patients. The delay times for the arterial, portal venous, and equilibrium phases were 30 s, 60 s, and 120 s respectively.

Image segmentation

The computer-assisted surgery system (CAS) (CAS-Lv, Qingdao Hisense Medical Equipment Co., Ltd.) was used to segment the liver contour automatically from the hepatobiliary phase after 15 minutes of MRI and venous phase of the CT, to obtain the regions of interest (ROI) of the liver. For liver segmentation of MRI and CT, the Dice coefficient was more than 0.95 (this Dice coefficient is the manufacturer reference data of CAS-Lv). Each automatically segmented liver contour was visually inspected and any inaccurate liver contour were manually corrected by a doctor with over a decade of experience. All patients’ images and liver contour were saved as .NII files.

Radiomic feature extraction

The preprocessing steps include Pixelspacing resampling and normalization of the image. We extracted 107 radiomics features from CT and MRI of patient. The features were split into seven different groups: (1) first-order statistics of voxel intensity features (n = 18), (2) shape features (n = 14), (3) gray level co-occurrence matrix (GLCM) features (n = 24), (4) gray level dependence matrix (GLDM) features (n = 14), (5) gray level run-length matrix(GLRLM) features (n = 16), (6) gray level size zone matrix (GLSZM) features (n = 16), and (7) neighboring gray tone difference matrix (NGTDM) features (n = 16). PyRadiomics package implemented in Python (version 3.7) was utilized to automatically conduct the feature extraction process. In addition, each part of the feature name was concatenated with underlines, it includes image type, feature group, and the feature name. For example, original_firstorder_Skewness represents a feature extracted from the original image, firstorder group, and feature name was Skewness.

Radiomic feature selection

At first, statistical tests were performed to calculate the p value between the features. When the two-tailed p value of features was p < 0.1, the features was retained. Second, spearman correlation analysis was performed to calculate the correlation coefficient between features. When the correlation coefficient of many features was r༞0.9, one feature randomly retained. This step was applied to reduce the collinearity of features. At last, the least absolute shrinkage and selection operator (LASSO) algorithm was used to reduce the unimportant features. The statistical tests, correlation analysis, and LASSO algorithm were also implemented by importing the “scipy”, “numpy”, and “sklearn” packages in Python (version 3.7).

Model construction and performance evaluation

Supervised learning was used for training and prediction. More specifically, six machine learning algorithms were applied to investigate the performance of the model, whereas these classifiers were K-Nearest Neighbors (KNN), Support Vector Machines (SVM), Extra Trees(ET), Random Forest (RF), Light Gradient Boosting Machine (LightGBM) and eXtreme Gradient Boosting (XGBoost). All selected features were used as input to classify the evaluation of functional liver reserve (ICG-R15 ≤ 10% vs ICG-R15༞10%, ICG-R15 ≤ 20% vs ICG-R15༞20%, and ICG-R15 ≤ 30% vs ICG-R15༞30% as 2-class classifier). All patients were randomly split into two cohorts. One was called the training dataset (80%) and the other was called the test dataset (20%). Each of the models was trained on the training dataset, and five cross-validated was performed on the dataset by using StratifiedKFold, which was provided by the “sklearn” package in Python (version 3.7). StratifiedKFold can ensure that each set contains approximately the percentage of samples in each class, which can reduce the sampling deviation of the set. We used the ACC and AUC of ROC with 95% CI to evaluate the performance of classification models. The models and values calculation was implemented by importing the “pandas” and “sklearn” packages in Python (version 3.7).

Feature selection

The features were selected by conducting statistical tests, spearman correlation analysis, and LASSO. Therefore, Table 2 shows the features and weight coefficients of Gd-EOB-DTPA-enhanced hepatic MRI and contrast-enhanced CT.

Table 2

Feature selection results and weights
Data	ICG-R15	Features (Training set)	Coefficient
MRI	ICG-R15 ≤ 10% vs ICG-R15༞10%	original_glszm_SmallAreaLowGrayLevelEmphasis	0.063048
		original_gldm_GrayLevelNonUniformity	0.038531
		original_glszm_ZoneVariance	0.031718
		original_firstorder_Skewness	0.012104
		original_glszm_SmallAreaEmphasis	-0.00618
		original_shape_Sphericity	-0.08739
	ICG-R15 ≤ 20% vs ICG-R15༞20%	original_shape_SurfaceVolumeRatio	0.126727
		original_glszm_LargeAreaHighGrayLevelEmphasis	0.077114
		original_firstorder_Kurtosis	0.065784
		original_glszm_SmallAreaLowGrayLevelEmphasis	0.008947
		original_glszm_ZoneVariance	0.007758
		original_glrlm_LongRunHighGrayLevelEmphasis	-0.04705
		original_firstorder_Minimum	-0.076842
	ICG-R15 ≤ 30% vs ICG-R15༞30%	original_firstorder_Kurtosis	0.102976
		original_glszm_LargeAreaHighGrayLevelEmphasis	0.070309
		original_glszm_SmallAreaLowGrayLevelEmphasis	0.048927
		original_shape_SurfaceVolumeRatio	0.029746
		original_glszm_SizeZoneNonUniformity	-0.00487
		original_shape_MinorAxisLength	-0.004987
		original_shape_LeastAxisLength	-0.024224
CT	ICG-R15 ≤ 10% vs ICG-R15༞10%	original_glszm_ZoneEntropy	0.146007
		original_glszm_SmallAreaEmphasis	-0.069739
		original_firstorder_RootMeanSquared	-0.082279
		original_firstorder_Energy	-0.14831
	ICG-R15 ≤ 20% vs ICG-R15༞20%	original_shape_SurfaceVolumeRatio	0.10414
		original_glcm_ClusterProminence	0.062316
		original_glszm_LargeAreaLowGrayLevelEmphasis	0.017001
		original_ngtdm_Strength	0.015159
		original_glszm_LowGrayLevelZoneEmphasis	0.005915
		original_glrlm_GrayLevelNonUniformityNormalized	-0.01747
		original_firstorder_Energy	-0.027848
		original_glszm_SmallAreaEmphasis	-0.040413
		original_firstorder_RootMeanSquared	-0.094183
	ICG-R15 ≤ 30% vs ICG-R15༞30%	original_firstorder_Kurtosis	0.102976
		original_glszm_LargeAreaHighGrayLevelEmphasis	0.070309
		original_glszm_SmallAreaLowGrayLevelEmphasis	0.048927
		original_shape_SurfaceVolumeRatio	0.029746
		original_glszm_SizeZoneNonUniformity	-0.00487
		original_shape_MinorAxisLength	-0.004987
		original_shape_LeastAxisLength	-0.024224

Predicted model

For ICG-R15 = 10%, ICG-R15 = 20%, and ICG-R15 = 30% as functional liver reserve thresholds, six machine learning algorithms were used to train the trainning dataset and predict the result on the test dataset. The detailed performance of six models is stated in Table 3. Among all of MRI groups of the result, more specifically, when ICG-R15 = 10% was used as a threshold, the classifier LightGBM achieved the highest performance, an AUC value of 0.932 (95% CI: 0.795–1), and ACC of 0.955 were derived. When ICG-R15 = 20% was used as a threshold, the classifier LightGBM achieved the highest performance, an AUC value of 0.938 (95% CI: 0.838–1), and ACC of 0.913 were extracted. When ICG-R15 = 30% was utilized as a threshold, the classifier XGBoost achieved the highest performance, an AUC value of 0.972 (95% CI: 0.909–1), and ACC of 0.955 were derived. By taking into account all the results of CT groups, when ICG-R15 = 10% was selected as a threshold, the classifier LightGBM achieved the highest performance with an AUC value of 0.891 (95% CI: 0.785–0.996), and ACC of 0.868. When ICG-R15 = 20% was selected as a threshold, the classifier SVM achieved the highest performance with an AUC value of 0.877 (95% CI: 0.766–0.988), and ACC of 0.842. When ICG-R15 = 30% was selected as a threshold, the classifier LightGBM achieved the highest performance with an AUC value of 0.927 (95% CI: 0.792-1), and ACC of 0.947. The detailed performance of the models is listed in Table 4. The model confusion matrix of the MRI and CT are also provided (Fig. 2). Furthermore, the best performance models ROC curve was derived (Fig. 3).

Table 3

Performance comparison of machine learning model
Data	ICG-R15		SVM	KNN	RF	ExtraTrees	XGBoost	LightGBM
MRI	ICG-R15 ≤ 10% vs ICG-R15༞10%	Test set ACC	0.909	0.864	0.864	0.727	0.773	0.955
		Test set AUC (95%CI)	0.949 (0.864-1)	0.876 (0.730-1)	0.850 (0.679-1)	0.778 (0.573–0.983)	0.803 (0.588-1)	0.932 (0.795-1)
	ICG-R15 ≤ 20% vs ICG-R15༞20%	Test set ACC	0.826	0.607	0.739	0.609	0.783	0.913
		Test set AUC (95%CI)	0.884 (0.747-1)	0.688 (0.470–0.905)	0.625 (0.336–0.914)	0.661 (0.444–0.878)	0.759 (0.542–0.976)	0.938 (0.838-1)
	ICG-R15 ≤ 30% vs ICG-R15༞30%	Test set ACC	0.864	0.818	0.909	0.864	0.955	0.864
		Test set AUC (95%CI)	0.917 (0.797-1)	0.778 (0.472-1)	0.944 (0.846-1)	0.847 (0.650-1)	0.972 (0.909-1)	0.931 (0.818-1)
CT	ICG-R15 ≤ 10% vs ICG-R15༞10%	Test set ACC	0.842	0.789	0.789	0.658	0.789	0.868
		Test set AUC (95%CI)	0.877 (0.767–0.988)	0.789 (0.654–0.943)	0.863 (0.652-1)	0.800 (0.657–0.943)	0.837 (0.693–0.981)	0.891 (0.785–0.996)
	ICG-R15 ≤ 20% vs ICG-R15༞20%	Test set ACC	0.842	0.553	0.684	0.711	0.579	0.711
		Test set AUC (95%CI)	0.877 (0.766–0.988)	0.651 (0.474-827)	0.588 (0.376-0.800)	0.740 (0.571–0.909)	0.560 (0.360–0.760)	0.648 (0.339-1)
	ICG-R15 ≤ 30% vs ICG-R15༞30%	Test set ACC	0.921	0.947	0.789	0.789	0.816	0.947

Table 4

Performance of the best MRI and CT machine learning classification model
Data	ICG-R15	ACC	AUC (95%CI)	Sensitivity	Specificity	PPV	NPV	FNR	Model
MRI	ICG-R15 ≤ 10% vs ICG-R15༞10%	0.955	0.932(0.795-1)	1	0.889	0.929	1	0	LightGBM
	ICG-R15 ≤ 20% vs ICG-R15༞20%	0.913	0.938(0.838-1)	1	0.875	0.778	1	0	LightGBM
	ICG-R15 ≤ 30% vs ICG-R15༞30%	0.955	0.972(0.909-1)	1	0.944	0.800	1	0	XGBoost
CT	ICG-R15 ≤ 10% vs ICG-R15༞10%	0.868	0.891(0.785–0.996)	0.960	0.692	0.857	0.900	0.04	LightGBM
	ICG-R15 ≤ 20% vs ICG-R15༞20%	0.842	0.877(0.766–0.988)	0.923	0.800	0.706	0.952	0.077	SVM
	ICG-R15 ≤ 30% vs ICG-R15༞30%	0.947	0.927(0.792-1)	0.833	0.969	0.833	0.969	0.167	LightGBM
ICG-R15: indocyanine green retention rate at 15 min, AUC: Area under the ROI curve, ACC: Accuracy, PPV: Positive predictive value, FNR: False negative rate, NPV: negative predictive value,

From the extracted outcomes, it can be argued that both MRI-based and CT-based machine learning models can achieve the goal of classification in distinguishing patients with abnormal functional liver reserve. The performance of the MRI model was also better than that of the CT model in the assessment of functional liver reserve. For example, the accuracy of the MRI model was high (ICG-R15 ≤ 10% vs ICG-R15༞10%:MRI 0.955, CT 0.868; ICG-R15 ≤ 20% vs ICG-R15༞20%:MRI 0.913, CT 0.842; ICG-R15 ≤ 30% vs. ICG-R15༞30%:MRI 0.955, CT 0.947), the AUC of the MRI model was high (ICG-R15 ≤ 10% vs ICG-R15༞10%:MRI 0.932, CT 0.891; ICG-R15 ≤ 20% vs ICG-R15༞20%:MRI 0.938, CT 0.877; ICG-R15 ≤ 30% vs ICG-R15༞30%:MRI 0.986, CT 0.927), and the false negative rate (FNR) was 0 (ICG-R15 ≤ 10% vs ICG-R15༞10%:MRI 0, CT 0.04;ICG-R15 ≤ 20% vs ICG-R15༞20%:MRI 0, CT 0.07;ICG-R15 ≤ 30% vs ICG-R15༞30%:MRI 0, CT 0.167).

Radiomics has shown great value in diagnosis and therapy of multiple diseases. Compared with ICG-R15, it was examined whether it is possible to perform an accurate assessment of functional liver reserve by using Gd-EOB-DTPA-enhanced hepatic MRI and contrast-enhanced CT in HCC patients based radiomics method. Under this perspective, MRI-based and CT-based machine learning models were developed and validated for distinguishing patients with functional liver reserves of different states. Although both models worked satisfactorily, the performance of the MRI model was better than that of the CT model for the assessment of functional liver reserve.

Recently, there has been an increasing application of imaging techniques in measuring the hepatic function in HCC patients, because it can provide more significant clinical information than an overall assessment [18]. The liver has regional differences in hepatic parenchymal abnormalities [19, 20]. One recent work from Zhaoqi Shi et al. showed that radiomics analysis can be applied in the preoperative assessment of functional liver reserve in HCC patients [21]. However, this research only focused on Gd-EOB-DTPA-enhanced hepatic MRI to predict the ICG classification value to evaluate liver function in HCC patients, and the classification of functional liver reserve threshold were set to ICG-R15 = 10%, ICG-R15 = 15%, and ICG-R15 = 20%. However, the analysis of the CT images and the comparison of MRI and CT images was not mentioned. In this work, not only contrast-enhanced CT images were used as another practicable evaluation reference to assess functional liver reserve, but also ICG-R15 = 10%, ICG-R15 = 20%, and ICG-R15 = 30% were set as functional liver reserve thresholds. By referring to the criteria of safe hepatic resection proposed by Makuuchi [22], if ICG-R15༜10%, trisegmentectomy and bisegmentectomy of the liver can be performed; If 10%≤ICG-R15༜20%, left lobictomy right monosegmentectomy of the liver can be performed; If 20%≤ICG-R15༜30%, subsegmentectomy of Couinaud of the liver can be performed; If ICG-R15 ≥ 30%, it is necessary to limit resection or enucleation liver for transplantation [22]. The different ICG-R15 values of the patient correspond to different surgical strategies. Here, the threshold of classification was increased to cover more ICG-R15 values in accordance with the mentioned Makuuchi criteria for safe hepatic resection, which can provide a better reference for clinical surgery. The contrast-enhanced CT is regarded as the most common modality for patients with HCC. Here, not only CT images were used as another practicable evaluation reference to assess functional liver reserve, but also CT was compared with MRI for evaluation. With the radiomics model, we could assist doctors to evaluate functional liver reserve in the hospital without expensive ICG equipment, and reduce the burden on patients.

Gd-EOB-DTPA- enhanced hepatic MRI and contrast-enhanced CT were selected as data for evaluating presurgical liver function. The hepatobiliary phase after 15 minutes was selected from (Gd-EOB-DTPA)-enhanced hepatic MRI for evaluating presurgical liver function data. Additionally, during hepatobiliary phase, HCCs appeared hypointense in Gd-EOB-DTPA-enhanced images because of contrast medium discharged into hepatocytes and bile ducts in hepatobiliary phase. Hence, diagnostic sensitivity and specificity of HCC were dramatically improved [23]. The obtained signals during Gd-EOB-DTPA-enhanced MRI imaging describe anatomical characteristics of liver and hepatocyte-specific function [23]. The effectiveness of Gd-EOB-DTPA-enhanced hepatic MRI for evaluating liver function has been evaluated by several works in the literature [24]. The portal venous phase was selected from contrast-enhanced CT for evaluating presurgical liver function data. Intense contrast uptake in arterial phase was conducted before extracellular contrast washout in portal venous and/or delayed phases [25]. During the portal venous phase, the CT value of HCCs was lower than liver parenchyma because of washout. We defined washout as a visually evaluated relative decrease in the enhancement to background liver on portal venous and/or delayed phase [26–29]. In our research, the features in different threshold groups were selected based on shape, voxel intensity, and texture to predict classification. The developed disaggregated model was too complex for clinical practice for it comprises more features. Therefore, further research is required to simplify the features. In this work, six classifiers were used for modeling. By comparing with other classifiers, a better performance was demonstrated for LightGBM. LightGBM is a gradient boosting framework on the basis of decision tree algorithms [30]. Classification tasks are often carried out with LightGBM [31]. Sensitivity to overfitting creates greater challenge to LightGBM algorithm, especially for small dataset [32]. Therefore, here we adjusted LightGBM model parameters to avoid overfitting.

However, there are several limitations that should be mentioned. First, the provided research images were obtained from our center, amount of patients was relatively little. External validation cohort was also not allocated. Besides, the baseline characteristics and features may not conform to population. Second, the threshold analysis of functional liver reserve is a 2-class classification problem, which cannot cover the all of the ICG-R15 value intervals with clinical significance mentioned in the Makuuchi criteria for safe hepatic resection [22]. In the future, the classification with massive patients and data from multi-center will be investigated in combination with the different ICG value intervals of multi-class classification problems to provide more clinical reference.

Both Gd-EOB-DTPA-enhanced hepatic MRI and contrast-enhanced CT machine learning models are valuable noninvasive methods for the assessment of functional liver reserve. Compared with the ICG Clearance Test, the prediction of the ML models has similar effectiveness. The performance of the MRI model was better than that of the CT model, in terms of assessing the functional liver reserve. Thereby, the possibility of patients with poor liver reserve being predicted as a better reserve can be significantly reduced. Our work provides valuable insights, which can help clinicians to develop personalized precision treatment strategies.

ACC: Accuracy; AST: aspartate aminotransferase; ALT: alanine transaminase; PT: prothrombin time; CT: computed tomography; CI: confidence intervals; MRI: magnetic resonance imaging; HCC: hepatocellular carcinoma; ET: Extra Trees; Gd-EOB-DTPA: Gadolinium ethoxybenzyl dimeglumine; ML: machine learning; ICG-R15: indocyanine green retention rate at 15 min; ROI: regions of interest; RF: Random Forest; AUC: Area under the ROI curve; ROC: receiver operating characteristic; NPV: negative predictive value; PPV: Positive predictive value; FNR: False negative rate; BMI: Body Mass Index; HBV: Hepatitis B Virus; ALB: albumin; TBIL: total Bilirubin; GGT: gamma-glutamyltransferase; SVM: Support Vector Machines; LASSO: least absolute shrinkage and selection operator；KNN: K-Nearest Neighbors; XGBoost: eXtreme Gradient Boosting; LightGBM: Light Gradient Boosting Machine.

Acknowledgements

We thank all members in Shandong Key Laboratory of Digital Medicine and Computer Assisted Surgery, Affiliated Hospital of Qingdao University, for constructive advice in manuscript preparation.

Funding

Our study was funded by Natural Science Foundation of Shandong Province (grant number ZR2021MH171), Shandong Higher Education Young Science and Technology Support Program (grant number 2020KJL005) and Taishan Scholars Program of Shandong Province (grant number 2019010668).

Ethics approval and consent to participate

This study was approved by the Ethics Committee of the Affiliated Hospital of Qingdao University. The approval number of the Ethics Review was QYFY-WZLL-27465. The need for informed consent was waived by the Ethics Committee of the Affiliated Hospital of Qingdao University due to its retrospective nature. The methods of this study were carried out in accordance with relevant guidelines and regulations.

Consent for publication

Not applicable.

Availability of data and materials

The data and materials in this study are available through the corresponding authors with permission of the Affiliated Hospital of Qingdao University.

Competing interests

ZL worked for Qingdao Hisense Medical Equipment Co., Ltd.

The authors declare that their study was performed without any financial or commercial relation that may be interpreted as potential conflict of interest.

Authors' contributions

LZ and CZ devoted themselves to conception and design. LZ, FW, JC and LZ organized the database, LZ wrote the draft of the manuscript, CZ contributed to review and revision. All authors mentioned made a substantial and intellectual contribution to this study directly and approved its publication.

Author details

¹Shandong Key Laboratory of Digital Medicine and Computer Assisted Surgery, Affiliated Hospital of Qingdao University, Qingdao, China. ²Department of Radiology, Affiliated Hospital of Qingdao University, Qingdao, China. ³Qingdao Hisense Medical Equipment Co., Ltd, Qingdao, China. ⁴Department of Hepatobiliary and Pancreatic Surgery, Affiliated Hospital of Qingdao University, Qingdao, China.

Sung H, Ferlay J, Siegel RL, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin. 2021;71(3):209–49.
GLOBOCAN. International Agency for Research on Cancer. Published 2020. Cited 2021. http://gco.iarc.fr/today/home.
Llovet JM, Zucman-Rossi J, Pikarsky E, et al. Hepatocellular carcinoma. Nat Rev Dis Primers. 2016;2:16018.
Fu J, Wang H. Precision diagnosis and treatment of liver cancer in China. Cancer Lett. 2018;412:283–88.
van den Broek MA, Olde Damink SW, Dejong CH, et al. Liver failure after partial hepatic resection: definition, pathophysiology, risk factors and treatment. Liver Int. 2008;28(6):767–80.
Li B, Qin Y, Qiu Z, Ji J, Jiang X. A cohort study of hepatectomy-related complications and prediction model for postoperative liver failure after major liver resection in 1,441 patients without obstructive jaundice. Ann Transl Med. 2021;9(4):305.
Takamoto T, Ban D, Nara S, et al. Automated Three-Dimensional Liver Reconstruction with Artificial Intelligence for Virtual Hepatectomy. J Gastrointest Surg. 2022;26(10):2119–27.
Greco E, Nanji S, Bromberg IL, et al. Predictors of peri-opertative morbidity and liver dysfunction after hepatic resection in patients with chronic liver disease. HPB (Oxford). 2011;13(8):559–65.
De Gasperi A, Mazza E, Prosperi M. Indocyanine green kinetics to assess liver function: Ready for a clinical dynamic assessment in major liver surgery? World J Hepatol. 2016;8(7):355–67.
Choi JY. Radiomics and Deep Learning in Clinical Imaging: What Should We Do? Nucl Med Mol Imaging. 2018;52(2):89–90.
Peeken JC, Nüsslin F, Combs SE. "Radio-oncomics": The potential of radiomics in radiation oncology. Strahlenther Onkol. 2017;193(10):767–79.
Vial A, Stirling D, Field M, Ros M, Ritz C, Carolan M, et al. The role of deep learning and radiomic feature extraction in cancer-specific predictive modelling: a review. Transl Cancer Res. 2018;7(3):803–16.
Cook TS. The Importance of Imaging Informatics and Informaticists in the Implementation of AI. Acad Radiol. 2020;27(1):113–16.
Lambin P, Leijenaar RTH, Deist TM, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol. 2017;14(12):749–62.
Wei J, Jiang H, Gu D, et al. Radiomics in liver diseases: Current progress and future opportunities. Liver Int. 2020;40(9):2050–63.
Chen C, Chen C, Ma M, et al. Classification of multi-differentiated liver cancer pathological images based on deep learning attention mechanism. BMC Med Inform Decis Mak. 2022;22(1):176.
Seyama Y, Kokudo N. Assessment of liver function for safe hepatic resection. Hepatol Res. 2009;39(2):107–16.
Geisel D, Lüdemann L, Hamm B, Denecke T. Imaging-Based Liver Function Tests–Past, Present and Future. Rofo. 2015;187(10):863–71.
Kwon AH, Matsui Y, Ha-Kawa SK, Kamiyama Y. Functional hepatic volume measured by technetium-99m-galactosyl-human serum albumin liver scintigraphy: comparison between hepatocyte volume and liver volume by computed tomography. Am J Gastroenterol. 2001;96(2):541–6.
Ratziu V, Charlotte F, Heurtier A, et al. Sampling variability of liver biopsy in nonalcoholic fatty liver disease. Gastroenterology. 2005;128(7):1898–906.
Shi Z, Cai W, Feng X, et al. Radiomics Analysis of Gd-EOB-DTPA Enhanced Hepatic MRI for Assessment of Functional Liver Reserve. Acad Radiol. 2022;29(2):213–8.
Makuuchi M, Kosuge T, Takayama T, et al. Surgery for small liver cancers. Semin Surg Oncol. 1993;9(4):298–304.
Van Beers BE, Pastor CM, Hussain HK. Primovist, Eovist: what to expect? J Hepatol. 2012;57(2):421–9.
Bae KE, Kim SY, Lee SS, et al. Assessment of hepatic function with Gd-EOB-DTPA-enhanced hepatic MRI. Dig Dis. 2012;30(6):617–22.
Ayuso C, Rimola J, Vilana R, et al. Diagnosis and staging of hepatocellular carcinoma (HCC): current guidelines [published correction appears in Eur J Radiol. 2019 Mar;112:229]. Eur J Radiol. 2018;101:72–81.
Elsayes KM, Hooker JC, Agrons MM, et al. 2017 Version of LI-RADS for CT and MRI Imaging: An Update. Radiographics. 2017;37(7):1994–2017.
Tang A, Bashir MRI, Corwin MT, et al. Evidence Supporting LI-RADS Major Features for CT- and MRI Imaging-based Diagnosis of Hepatocellular Carcinoma: A Systematic Review. Radiology. 2018;286(1):29–48.
Choi JY, Lee JM, Sirlin CB. CT and MRI imaging diagnosis and staging of hepatocellular carcinoma: part II. Extracellular agents, hepatobiliary agents, and ancillary imaging features. Radiology. 2014;273(1):30–50.
Yoon JH, Park JW, Lee JM. Noninvasive Diagnosis of Hepatocellular Carcinoma: Elaboration on Korean Liver Cancer Study Group-National Cancer Center Korea Practice Guidelines Compared with Other Guidelines and Remaining Issues. Korean J Radiol. 2016;17(1):7–24.
Ke G, Meng Q, Finley T, et al. LightGBM: A highly efficient gradient boosting decision tree. NIPS. 2017;3146–54.
Zeng H, Yang C, Zhang H, et al. A LightGBM-Based EEG Analysis Method for Driver Mental States Classification. Comput Intell Neurosci. 2019;2019(4):1–11.
Zhao Q, Ye Z, Su Y, Ouyang D. Predicting complexation performance between cyclodextrins and guest molecules by integrated machine learning and molecular modeling techniques. Acta Pharm Sin B. 2019;9(6):1241–52.

No competing interests reported.

Download PDF

Journal Publication

published 17 Jul, 2023

Read the published version in BMC Medical Imaging →

Editorial decision: Major revision
06 Apr, 2023
Reviews received at journal
16 Jan, 2023
Reviewers agreed at journal
16 Jan, 2023
Reviewers agreed at journal
02 Jan, 2023
Reviewers invited by journal
18 Dec, 2022
Editor assigned by journal
17 Dec, 2022
Editor invited by journal
02 Dec, 2022
Submission checks completed at journal
02 Dec, 2022
First submitted to journal
24 Nov, 2022

You are reading this latest preprint version

Machine learning-based Radiomics analysis of preoperative functional liver reserve with MRI and CT image

Status:

Journal Publication

Version 1

Abstract

Figures

Introduction

Methods

Results

Discussion

Conclusion

Abbreviations

Declarations

Acknowledgements

Funding

Ethics approval and consent to participate

Consent for publication

Availability of data and materials

Competing interests

Authors' contributions

Author details

References

Additional Declarations

Status:

Journal Publication

Version 1