Enter the Era of Quantitative Liver CT.

Abstract

HomeRadiologyVol. 289, No. 3 PreviousNext Reviews and CommentaryFree AccessEditorialEnter the Era of Quantitative Liver CTAndrew D. Smith Andrew D. Smith Author AffiliationsFrom the Department of Radiology, University of Alabama, 619 19th St South, JTN 452, Birmingham AL 35249-6830.Address correspondence to the author (e-mail: [email protected]).Andrew D. Smith Published Online:Sep 4 2018https://doi.org/10.1148/radiol.2018181847MoreSectionsPDF ToolsImage ViewerAdd to favoritesCiteTrack CitationsPermissionsReprints ShareShare onFacebookTwitterLinked InEmail See also the articles by Sartoris et al and Choi et al in this issue.IntroductionThe era of quantitative liver CT is upon us, with image-processing algorithms that are linked to meaningful clinical outcomes. The driving factors behind this movement are the widespread use and relative standardization of liver CT across different platforms and continued advancement of image-processing algorithms.In this issue of Radiology, the articles by Sartoris et al (1) and Choi et al (2) explore new methods for evaluating chronic liver disease and cirrhosis by using quantitative information derived from liver CT images. Both author groups demonstrated high accuracy using routine CT images, suggesting that the techniques may be widely applicable. Similarly, both groups used existing liver CT images, which add no additional patient expense or ionizing radiation, and this indicates that subsequent validation could be conducted with retrospective studies, which are less expensive and more rapid to complete than prospective studies.Sartoris et al (1) conducted a retrospective study to determine the accuracy of the CT-based liver surface nodularity (LSN) score to estimate clinically significant portal hypertension in patients with cirrhosis. The current reference standard for evaluating portal pressure is to measure the hepatic venous pressure gradient, but this technique is invasive, difficult to reproduce, and not widely available. Noninvasive methods for assessing portal pressure in cirrhosis include assessment of liver stiffness at transient elastography, but this technique must be performed prospectively, is uncommonly performed in patients with cirrhosis, and is limited by technical challenges.The underlying hypothesis is that cirrhotic nodules progressively increase in size and number in cirrhosis, are visible at CT at the liver surface, and are associated with clinically significant portal hypertension. The LSN score is fundamentally a length measurement derived from CT images. More specifically, the LSN score is the average distance between each pixel of the detected surface of the liver and a mathematically smoothed line derived from the detected surface that is designed to mimic a normal smooth liver surface. The authors measured LSN on an average of 10 sections by using previously validated semiautomated software with an average processing time of less than 2 minutes.Sartoris et al (1) compared the CT-based LSN score to multiple serum laboratory indexes and to liver and splenic volumetric measurements. In a training cohort (n = 203), the LSN score had technical failures in 7% (14 of 203). In the remaining patients (n = 189), the LSN score correlated strongly with hepatic venous pressure gradient measurements (r = 0.75), had high accuracy (area under the receiver operating characteristic curve, 0.88) for estimation of clinically significant portal hypertension, and was statistically associated with the grade of esophageal varices. The accuracy of the LSN score was significantly higher than multiple serum laboratory indexes and liver and splenic volumetric measurements. The authors used an independent validation cohort (n = 84) and again showed that the LSN score strongly correlated with hepatic venous pressure gradient measurements (r = 0.79) and had high accuracy (area under the receiver operating characteristic curve, 0.87) for estimation of clinically significant portal hypertension.The strengths of the study design include an excellent reference standard (ie, hepatic venous pressure gradient) obtained within 60 days of liver CT, a large patient cohort, independent training and validation cohorts, and a comprehensive analysis compared with multiple other noninvasive methods. The LSN score has been previously validated for staging chronic liver disease and cirrhosis (3–5), but to our knowledge this is the first study directly linking the LSN score to portal hypertension, the main driver of most complications of cirrhosis, including gastroesophageal variceal bleeding. The main limitations are the retrospective single-institution design and absence of a direct comparison to liver stiffness techniques, which were not available in most patients. In addition, all of the CT examinations were performed by using scanners manufactured by GE Healthcare, which limits generalizability. Of note, a recent study (6) that used a phantom containing simulated smooth and nodular liver surfaces demonstrated high repeatability across a wide range of CT image acquisition parameters and high reproducibility across 22 different scanners from different manufacturers, which suggested that the results from the study by Sartoris et al (1) may translate to CT scanners from other manufacturers.Choi et al (2) aimed to develop a deep learning system for staging liver fibrosis by using routine portal venous phase liver CT images. The current reference standard for staging liver fibrosis is a liver biopsy with pathologic staging, but this technique is invasive and prone to sampling error and subjectivity at pathologic staging. Current noninvasive methods for staging liver fibrosis include measurements of liver stiffness at US and MR elastography. The limitations of these methods include technically challenging examinations with occasional failed examinations, a need for expensive dedicated hardware, a lack of standardization across different platforms (especially for US elastography), and a lack of agreement for appropriate cut-off values.Choi et al (2) describe the development of a fully automated method to segment the liver from the CT images and then a separate fully automated deep learning system to extract image features and derive the stage of liver fibrosis. Such an algorithm would be widely applicable and could replace the need for a biopsy and provide an alternative to liver stiffness techniques. A fully automated algorithm could be applied retrospectively or prospectively, and theoretically would improve standardization compared with manual methods.Choi et al (2) used a large development data set (n = 7461) with staging of liver fibrosis by biopsy that was obtained within 3 months of contrast agent–enhanced portal venous phase liver CT. This is appropriate; however, their development data set was heterogeneous and included patients who underwent liver mass resection, liver transplantation for cirrhosis or hepatocellular carcinoma, or percutaneous liver biopsy for living liver donor work-up, notably not including patients with chronic liver disease who present for a random liver biopsy to stage liver fibrosis. Furthermore, the development data set was imbalanced: 45.0% of patients (3357 of 7461) had no liver fibrosis (stage F0), 43.5% of patients (3247 of 7461) had cirrhosis (stage F4, many had end-stage cirrhosis), and only 11.5% of patients (857 of 7461) had intermediate stages of liver fibrosis (stage F1–F3). To overcome the imbalance, the authors augmented the group of patients with intermediate stages of liver fibrosis by rotating the CT images between −15° and 15°, and/or by adding random Gaussian noise on a random basis. This type of augmentation is a way to “trick” the deep learning algorithm into viewing these modified images as though they are from unique patients. This is not the same as having truly unique patients, but it does mediate some of the imbalance.After initial development of the deep learning algorithm, Choi et al (2) evaluated the accuracy of the algorithm in three independent test data sets (891 total). Test data set 1 (n = 421) is the most clinically relevant because it included patients who underwent a liver biopsy for evaluation of abnormal liver function tests or for suspected chronic liver disease. Test data set 2 (n = 298) mostly included patients who underwent liver resection, and test data set 3 (n = 172) included patients with hepatic tumor or tumors and selected patients who underwent liver biopsy for elevated liver function tests or for suspected or known chronic liver disease. The accuracy for the deep learning algorithm for differentiating F0–F1 from F2–F4, F0–F2 from F3–F4, and F0–F3 from F4 liver fibrosis stages was exceptional with all test data sets. More importantly, the accuracy with test data set 1 was high (area under the receiver operating characteristic curve range, 0.94–0.97), exceeding the accuracy of serum fibrosis tests (area under the receiver operating characteristic curve range, 0.65–0.85) and the accuracy of four independent radiologists subjectively evaluating liver morphologic findings (area under the receiver operating characteristic curve range, 0.74–0.88). This accuracy far exceeds that from a comparative deep learning algorithm for staging liver fibrosis at CT (area under the receiver operating characteristic curve range, 0.73–0.76) in a smaller patient cohort (7).The strengths of the study design include a large sample size, separate independent test data sets, and comparisons to established serum laboratory indexes and subjective visual interpretations by four independent radiologists. The fully automated algorithm is of substantial interest because it has the potential to improve reproducibility compared with manual or semiautomated methods. The main limitation is the potential for overfitting of the deep learning algorithm to the development data set. The development data set included a high rate of chronic hepatitis B virus infection (40%), which is uncommon in Western populations. Finally, it is unclear how the deep learning algorithm classifies patients, and this so-called black-box issue is particularly important when applying this algorithm to new situations such as low radiation dose acquisition protocols, dual-energy CT, or newer image reconstruction algorithms. It may be difficult for regulatory bodies, payers, clinicians, and patients to base clinical treatments on an unknown mechanism. Further external validation in different patient populations will be critical.In conclusion, quantitative liver CT image-processing algorithms are on the rise, are widely applicable, and can be applied retrospectively to routine liver CT images. Sartoris et al (1) demonstrated high accuracy of the CT-based LSN score for helping to predict clinically significant portal hypertension. The semiautomated processing is rapid and can be applied to unenhanced or contrast-enhanced CT images, and the mechanism of action of the LSN score is well defined. By comparison, Choi et al developed a new deep learning algorithm on the basis of portal venous phase CT images that accurately staged liver fibrosis in patients with chronic liver disease. The fully automated processing is of considerable interest and may facilitate rapid validation in other patient populations, which is critical because the mechanism of action is not defined.Disclosures of Conflicts of Interest: A.D.S. Activities related to the present article: disclosed no relevant relationships. Activities not related to the present article: disclosed money to author’s institution for grant from NIH and RSNA medical student grant, and both grants are related to the liver surface nodularity score; disclosed money to author for patents received and pending related to the liver surface nodularity score and Liver Nodularity; disclosed that author is the CEO and owner of Liver Nodularity. Other relationships: disclosed no relevant relationships.