Multiphase CT scans findings for AI training in the differentiation of hepatocellular carcinoma and intrahepatic cholangiocarcinoma based on interobserver agreement of expert abdominal radiologist.

doi:10.21203/rs.3.rs-2953921/v1

Download PDF

Research Article

Multiphase CT scans findings for AI training in the differentiation of hepatocellular carcinoma and intrahepatic cholangiocarcinoma based on interobserver agreement of expert abdominal radiologist.

https://doi.org/10.21203/rs.3.rs-2953921/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Computed tomography (CT) is the imaging modality used to evaluate liver nodule and differentiate hepatocellular carcinoma (HCC) from intrahepatic cholangiocarcinoma (ICC). The purpose of this study was to find out potential CT features in the differentiation of hepatocellular carcinoma and intrahepatic cholangiocarcinoma. Patients with radiological and pathologically confirmed diagnosis of HCC and ICC between January 2013 and December 2015 were included. Two board certified diagnostic radiologists independently reviewed multiphase CT images. Most of HCC showed arterial hyperenhancement at 95.8%. Arterial enhancement in ICC was less often at 15.4% to 26.9%. Two readers showed a moderate agreement of cirrhosis appearance in both HCC and ICC. Cirrhosis appeared in HCC group more frequently than ICC group. Lymph node enlargement was more common seen in ICC than HCC and agreement between readers was almost perfect. Portal vein invasion in HCC was seen 14.6 % by both readers with a substantial agreement. We believed cirrhosis appearance and lymph node enlargement could be remarkable imaging feature in differentiated HCC and CCA.

Hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma (ICC) are the most common primary liver cancers, accounting for approximately 85% (1). Underlying chronic hepatic inflammation with viral hepatitis B or C is a risk factor for HCC, but less for ICC (1). The correct diagnosis of HCC and ICC is crucial because the treatment of the two entities is different.

Computed tomography (CT) is the imaging modality used to evaluate liver nodules and differentiate HCC from ICC. Because these two liver cancers have different enhancing patterns on multiphase CT. The typical HCC shows arterial hyperenhancement, washout appearance in the portovenous phase, and presence of enhancing capsule. Vascular invasion, particularly portal vein, is an imaging finding of advanced HCC. The American College of Radiology proposed the Liver Imaging-Reporting and Data System (LI-RADS) for standardized imaging diagnosis criteria in patients at risk of HCC (2). The LI-RADS 5 lesions are diagnostic of HCC without the need for histologic confirmation, Figure 1. In contrast to HCC, typical ICC shows rim-like enhancement at the tumor periphery during the arterial phase and progressive centripetal enhancement in a later phase, Figure 2. Intrahepatic bile duct dilatation is another radiological finding of ICC.

However, the atypical enhancement pattern of HCC and ICC has often been seen in daily practice, making a diagnosis of HCC or ICC is challenging Figure 3. The liver nodule in the high-risk patient might have an enhancing pattern similar to ICC. On other hand, small ICCs had enhancing patterns mimicking HCC, Figure 4. One study showed the accuracy of CT for distinguishing non-HCC malignancies from HCC was 79.9% (3). So, the diagnosis and differential diagnosis of HCC and ICC are sometimes difficult and require much experience from radiologists.

Artificial intelligence (AI) was developed many years ago. Machine learning (ML) and deep learning (DL) have been used in multiple studies in the field of medicine. The multi-layered neural network algorithms, convolutional neural network (CNN), have proven of great use in the analysis of radiology images (4, 5). Two studies have investigated the usefulness of DL-based liver tumor classification from a multiphase contrast-enhanced CT scan (6, 7).

Hereby, we try to find out the frequency and reliability of CT features that are commonly used by radiologists for the diagnosis of HCC and ICC. Then we could use these CT features for training the AI dataset for differential diagnosis HCC and ICC. Finally, we will improve the accuracy of AI applications in the field of hepatology.

This retrospective study was approved by the Research Ethics Committee of Faculty of Medicine, Chiang Mai Univeristy, Study Code: RAD-2562-06135 and waived the need for informed consent. Our study was performed in accordance with relevant guidelines and regulations.

Patient Selection

Patients with radiological and pathologically confirmed diagnoses of HCC and ICC between January 2013 and December 2015 were included in this retrospective study. The patient selection flow chart is presented in Figure 5. The inclusion criteria were patients with a diagnosis of HCC and ICC in hospital database, patients with pre-diagnosis multiphase CT scan in a picture archiving and communication system (PACS), and pathologically confirmed diagnosis of HCC or ICC in the Pathology Department database. The exclusion criteria were poor images quality or improper phases of CT scan, the pathological diagnosis of metastasis or other malignancies, and CT or pathological report perihilar or extrahepatic cholangiocarcinoma.

Imaging Acquisition

CT images of the liver were obtained with one of two multidetector CT scanners (SOMATOM FORCE and SOMATOM DEFINITION, Siemens Medical Solutions, Forchheim, Germany). The technique included pre-contrast imaging and multiphase post-contrast imaging following injection of 100 mL non-ionic iodinated contrast media (350 mg/mL) through the antecubital vein. The hepatic arterial phase (HAP), portovenous phase (PVP), and delayed phase (DP) of scanning began 35, 80, and 180 seconds after the injection of contrast media, respectively.

Imaging Interpretation

The CT image was independently reviewed on PACS and using Synapse workstation, by two expert board-certified Diagnostic Radiologists. Readers were blinded to the radiology and pathology reports. Each reader measured the tumor size on the most visible imaging phase. Each tumor was evaluated in terms of arterial hyperenhancement, portal vein invasion, and final diagnosis. The presence of cirrhosis and lymph node metastasis were evaluated

Histopathologic Examination

The histopathologic report was the reference standard of this study. Biopsy or surgical specimens of the hepatic nodule were fixed with 10% formalin and embedded in paraffin. The tissue slices were stained with hematoxylin-eosin and evaluated by pathologists. No reexamination of pathological slices was done.

Statistics

Statistical analyses were performed by using the software SPSS ver.19.

The quantitative data were presented as mean and standard deviation or median and interquartile range depending on data distribution. Categoric variables were analyzed by the chi-square test. The arterial hyperenhancement, portal vein thrombosis, lymph node enlargement, and cirrhosis appearance were calculated and presented to the sensitivity, specificity, likelihood ratio for diagnosis of HCC and ICC.

Inter-observed agreement of categorical data was evaluated using Cohen‘s kappa statistic. According to Landis and Koch, values less than 0.00 as indicating no agreement, 0.01-0.20 as slight agreement, 0.21-0.40 as fair agreement, 0.41-0.60 as moderate agreement, 0.61-0.80 as substantial agreement, and 0.81-1.00 as almost perfect agreement(8). Significance was defined as P < 0.05.

A total of 74 patients with a pathologically confirmed diagnosis, including 48 HCCs and 26 ICC, were included in this study. The mean age of patients with HCC was 52 years (range 21-76), while ICC was 57 years (range 35-76).

CT features of HCC and ICC and interobserver agreement

Frequent CT features and interobserver agreement between two readers were demonstrated in Tables 2 and 3.

Arterial enhancement in HCC was highly seen, 95.8% by both readers. Agreement between readers was moderate (k = 0.47). Arterial enhancement in ICC was less often, 15.4% by Reader 1 and 26.9% by Reader 2. Agreement between readers was substantial (k = 0.66).

Two readers showed a moderate agreement of cirrhosis appearance in both HCC and ICC. Cirrhosis appeared in HCC group more frequently than ICC group; 70.8% - 72.9% vs 42.3% - 61.5%)

Lymph node enlargement in HCC was not commonly seen, 14.6% by Reader 1 and 25.0% by Reader 2. Agreement between reader was fair (k = 0.29). Lymph node enlargement in ICC was more commonly seen, 50.0% by reader 1 and 57.7% by Reader 2. The agreement between readers was almost perfect (k = 0.84).

Portal vein invasion in HCC was seen 14.6 % by both readers with a substantial agreement (k= 0.66). Portal vein invasion in ICC was seen 11.5% by Reader 1 and 19.2% by Reader 2. Agreement both readers was substantial agreement (k= 0.70).

Correct Diagnosis

The diagnostic performance of the two radiologists was satisfactory with a corrected diagnosis of 87.8% and 94.6%, respectively. Both radiologists shared a common correct diagnosis of 65 out 74 patients, including 46 HCCs and 19 ICCs. Two HCCs and 7 ICCs, which were 4.15% of HCCs and 26.9% of ICCs, were misdiagnosis by both radiologists. The diagnostic performance of two expert radiologists in the diagnosis of HCC and ICC was summarized in Table 3

According to LI-RADS, the percentage of HCC detected by LR-5 criteria is 95%, while the percentage of malignancy is 98%(2). Most of HCC showed arterial hyperenhancement at 95.8% and interobserver agreement was moderate. The presence of arterial enhancement in our study was higher than the prior studies, probably due to HCCs in our study being bigger than other studies. The presence of arterial enhancement was related to arterial blood flow increased in the early stage of hepatocarcinogenesis in well-differentiated and moderately differentiated HCC. While poorly differentiated HCCs had decreased arterial blood flow, resulting in a typical enhancement pattern of some HCCs.

On the other hand, some ICC showed arterial enhancement. Using the presence of arterial hyperenhancement alone may lead to misdiagnosis. Interobserver agreement of presence or absence of arterial enhancement was substantial.

Two expert radiologists had high sensitivity in diagnosed HCCs with a high positive likelihood ratio and substantial agreement. In addition, ICCs were diagnosed with lower sensitivity, but higher specificity and positive likelihood ratio.

Cirrhosis can be associated with increase in HCC risk and ICC risk. Cirrhosis appearance has been used as a radiological finding for prediction of HCC. Our study demonstrated cirrhotic appearance of patient of HCCs more than patient of ICCs, around 70% versus 40 to 60%. Interobserver agreement of cirrhotic appearance was moderate. We preferred diagnosis of HCC in patients with cirrhotic appearance.

Lymph node enlargement or lymph node metastasis is commonly seen in ICC than HCC. Our study revealed lymph node metastasis in ICC patients about 50% with almost perfect agreement between two expert radiologists. We almost diagnosed ICCs when CT showed lymph node enlargement.

Machine learning methods were widely adopted in classifying liver cancer. Supervised learning with a set of data labeled in the algorithm such as support vector machine (SVM) has been used in differentiated liver cancer. Ponnoprat et al. obtained 88% in classification accuracy of differentiated HCC and ICC based on multiphase CT scan (6). Deep learning is currently studied in the classification of liver cancer. There are still several limitations due to selection bias of training dataset, small sample size, lack of internal and external validation. A convolutional neural network (CNN) is a kind of deep learning method that does not require the definition of specific radiological features to learn. It interprets a huge of training images and may discover additional differential features not yet identified in current radiological practice. Yasaka used CNN algorithm for differentiation of liver masses into five classes and found that the median area under the receiver operating characteristic curve for differentiating malignant and benign lesions was 0.92 (9).

Currently, Wang et al. proposed a proof-of-concept interpretable deep learning system for clinical radiology by identifying and scoring radiological features (10–12). He used 14 radiological features such as arterial phase hyperenhancement, central scar, enhancing rim, progressive centripetal filling, progressive hyperenhancement, thin-walled mass, and washout pattern to classify liver lesion into six classes: benign cyst, cavernous hemangioma, focal nodular hyperplasia, HCC, ICC, and metastasis. His model achieved 76.5% positive predictive value and 82.9% sensitivity in identifying the correct radiological features (11).

Our study has two limitations. First, this was a retrospective case-control study and may be subject to a selection bias. Second, in our study population, the ratio between HCC and ICC was different from the actual cases.

Our study focused on differentiated HCC and ICC. We tried to find out radiological features that may be useful in differential two entities. Our study found cirrhosis appearance was more commonly seen in the HCC group while lymph node metastasis was found in the ICC group. We believed these two radiological features could be ancillary and adopted in future training algorithms.

GRANT:

Supported by Faculty of Medicine, Chiang Mai University, grant no 150/2562

ETHICS APPROVAL:

This retrospective study was approved by the Research Ethics Committee, Faculty of Medicine, Chiang Mai University (Study Code: RAD-2562-06135), and informed consent was not required.

ACKNOWLEDGMENTS:

The authors thank Ruth Leatherman, RN of the Research Administration Section, Faculty of Medicine, Chiang Mai University for native English Proofreader.

DATA AVAILABILITY:

The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.

DISCLOSURE: No conflict of interest declared.

Massarweh NN, El-Serag HB. Epidemiology of Hepatocellular Carcinoma and Intrahepatic Cholangiocarcinoma. Cancer Control. 2017;24(3):1073274817729245.
Marrero JA, Kulik LM, Sirlin CB, Zhu AX, Finn RS, Abecassis MM, et al. Diagnosis, Staging, and Management of Hepatocellular Carcinoma: 2018 Practice Guidance by the American Association for the Study of Liver Diseases. Hepatology. 2018;68(2):723-50.
Ludwig DR, Fraum TJ, Cannella R, Ballard DH, Tsai R, Naeem M, et al. Hepatocellular carcinoma (HCC) versus non-HCC: accuracy and reliability of Liver Imaging Reporting and Data System v2018. Abdom Radiol (NY). 2019;44(6):2116-32.
Jimenez Perez M, Grande RG. Application of artificial intelligence in the diagnosis and treatment of hepatocellular carcinoma: A review. World J Gastroenterol. 2020;26(37):5617-28.
Yang YJ, Bang CS. Application of artificial intelligence in gastroenterology. World J Gastroenterol. 2019;25(14):1666-83.
Ponnoprat D, Inkeaw P, Chaijaruwanich J, Traisathit P, Sripan P, Inmutto N, et al. Classification of hepatocellular carcinoma and intrahepatic cholangiocarcinoma based on multi-phase CT scans. Med Biol Eng Comput. 2020;58(10):2497-515.
Shi W, Kuang S, Cao S, Hu B, Xie S, Chen S, et al. Deep learning assisted differentiation of hepatocellular carcinoma from focal liver lesions: choice of four-phase and three-phase CT imaging protocol. Abdom Radiol (NY). 2020;45(9):2688-97.
Landis JR, Koch GG. The measurement of observer agreement for categorical data. Biometrics. 1977;33(1):159-74.
Yasaka K, Akai H, Abe O, Kiryu S. Deep Learning with Convolutional Neural Network for Differentiation of Liver Masses at Dynamic Contrast-enhanced CT: A Preliminary Study. Radiology. 2018;286(3):887-96.
Hamm CA, Wang CJ, Savic LJ, Ferrante M, Schobert I, Schlachter T, et al. Deep learning for liver tumor diagnosis part I: development of a convolutional neural network classifier for multi-phasic MRI. Eur Radiol. 2019;29(7):3338-47.
Wang CJ, Hamm CA, Savic LJ, Ferrante M, Schobert I, Schlachter T, et al. Deep learning for liver tumor diagnosis part II: convolutional neural network interpretation using radiologic imaging features. Eur Radiol. 2019;29(7):3348-57.
Oestmann PM, Wang CJ, Savic LJ, Hamm CA, Stark S, Schobert I, et al. Deep learning-assisted differentiation of pathologically proven atypical and typical hepatocellular carcinoma (HCC) versus non-HCC on contrast-enhanced MRI of the liver. Eur Radiol. 2021;31(7):4981-90.

Table 1: Frequency of computed tomography (CT) features of histopathologically proven hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma (ICC) detected by expert radiologists.

	HCC (n=48)	ICC (n=26)
CT features	Frequency	Frequency
Arterial Enhancement
Reader 1	95.8%	15.4%
Reader 2	95.8%	26.9%

Cirrhosis appearance
Reader 1	70.8%	42.3%
Reader 2	72.9%	61.5%

Lymph node metastasis
Reader 1	14.6%	50.0%
Reader 2	25.0%	57.7%

Portal vein invasion
Reader 1	14.6%	11.5%
Reader 2	14.6%	19.2%

Table 2: Interobserver agreement among expert radiologist in computed tomography (CT) feature of histopathologically proven hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma (ICC)

	HCC		ICC
CT features	Kappa	Agreement	Kappa	Agreement
Arterial Enhancement	0.47	Moderate	0.66	Substantial
Cirrhosis appearance	0.43	Moderate	0.48	Moderate
Lymph node metastasis	0.29	Fair	0.84	Almost perfect
Portal vein invasion	0.66	Substantial	0.70	Substantial
Final Diagnosis	0.65	Substantial	0.52	Moderate

Table 3: Diagnostic performance of two expert radiologist in diagnosis of hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma (ICC)

	HCC				ICC
Diagnosis	sensitivity	spec	Positive Likelihood ratio	Negative Likelihood ratio	sensitivity	spec	Positive Likelihood ratio	Negative Likelihood ratio
Reader 1	.979	0.885	8.513043	0.024	.885	0.979	42.142	0.117
Reader 2	.958	0.731	3.561338	0.057	.731	0.958	17.404	0.280

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Multiphase CT scans findings for AI training in the differentiation of hepatocellular carcinoma and intrahepatic cholangiocarcinoma based on interobserver agreement of expert abdominal radiologist.

Status:

Version 1

Abstract

Figures

Introduction

Methods

Results

Discussion

Declarations

References

Tables

Additional Declarations

Status:

Version 1