Risk Stratification of Hepatocarcinogenesis Using a Deep Learning Based Clinical, Biological and Ultrasound Model in High-risk Patients

Summary

Participation Requirements

Description

By 2030, hepatocellular carcinoma (HCC) will become the second leading cause of cancer-related death, accounting for more than one million deaths per year according to the World Health Organization. To this date, screening for hepatocellular carcinoma in France remains uniform for all patients, base...

By 2030, hepatocellular carcinoma (HCC) will become the second leading cause of cancer-related death, accounting for more than one million deaths per year according to the World Health Organization. To this date, screening for hepatocellular carcinoma in France remains uniform for all patients, based solely on a liver ultrasound every 6 months. This scheme has the advantage of associating an acceptable cost-effectiveness ratio and, above all, of obtaining an increased overall survival. However, this strategy has three main limitations: lack of personalisation, low compliance, relatively poor performance of the ultrasound. Risk stratification models have been developed for chronic hepatitis C, alcoholic cirrhosis and non-alcoholic steatohepatitis (NASH) including clinical (age, sex, body mass index and diabetes) and biological (ASAT/ALAT, platelets, albumin) parameters. However, they didn't include analysis of the liver parenchyma which is the physiopathological substrate of hepatocarcinogenesis. In the 1990s, several authors studied the incidence of hepatocellular carcinoma according to the liver echostructure. They agreed on the over-risk represented by a nodular heterogeneous echostructure with an estimated rate ratio of up to 20. However, all these results have not yet led to a personalised radiological screening strategy. The advent of new artificial intelligence techniques could revolutionize the approach. Deep learning, a subclass of machine learning, is a popular area of research that can help humans performing certain tasks. Unlike radiomics, deep learning can automatically identify new image features not defined by humans. The hypothesis of this study is that the non-tumor cirrhotic liver parenchyma is rich in structural information reflecting the severity of the hepatopathy, its carcinological risk and the process of hepatocarcinogenesis. Its analysis combined with clinical and biological data, which have already been studied to stratify the risk of hepatocarcinogenesis, will allow to define a very high-risk population, particularly in the context of Hepatitis C Virus (HCV) eradication and Hepatitis B Virus (HBV) control. Consequently, this study proposes to design prospectively a deep learning model for stratification of the risk of hepatocarcinogenesis by including clinical, biological and radiological ultrasound parameters. The primary objective of the study is to identify a population at very high risk of developing hepatocarcinoma in order to propose different screening modalities to the patients most at risk. This clinical study will include patients aged over 18 years referred by their hepatologist in the framework of ultrasound screening according to the European Association for the Study of the Liver (EASL) recommendations for hepatocellular carcinoma screening, except for non-cirrhotic HBV liver disease: non-cirrhotic F3-stage liver disease from any cause based on individual risk assessment for hepatocarcinoma; cirrhosis from any cause, non-viral or virologically cured (HCV) or controlled (HBV). Patients with a history of treated hepatocellular carcinoma will be excluded. Two groups of patients will be constituted prospectively: group 1 will include patients with a diagnosis of hepatocellular carcinoma greater than 1 cm (reference diagnostic standards: radiological or histological). These patients will therefore correspond to a very high-risk; Group 2 will include patients without hepatocellular carcinoma, thus corresponding to a lower risk. A 1 year-interval ultrasound will be performed in patients of group 2 to confirm the absence of new nodule in the year following inclusion. The proportion of new hepatocellular carcinoma should not exceed 3%. The data collected will be clinical, biological, elastographic and ultrasonic parameters. A Deep Learning model using a deep convolutional neural network architecture will be developed on Python using these data. On a total of 7 investigation sites, 300 patients (equitably distributed between the two groups) will be included in the training/validation cohort and 100 patients (equitably distributed between the two groups) in the test cohort. These numbers are calculated from ultrasound studies reporting a rate ratio of HCC risk of up to 20 in case of macronodular ultrasound pattern and Deep Learning requirements (large numbers needed). The training/validation and test cohorts will be from external and independent centres. The diagnostic performance of the model will be estimated by Area Under the Curve (AUC), sensitivity, specificity and F1-score (95% confidence intervals) on the test cohort.

Tracking Information