Using a single abdominal computed tomography image to differentiate five contrast-enhancement phases: A machine-learning algorithm for radiomics-based precision medicine

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: Contrast media; Liver neoplasms; Machine learning; Quality control; Radiomics.

Fig. 1.

Flow chart.

Fig. 2.

Flow diagram.

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.

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).

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).