Using an AI model that combines patient images and prognostic biomarker information could improve early diagnosis of hepatocellular carcinoma (HCC) among cirrhosis patients.
The findings of abstract No. 340, “Artificial Intelligence-mediated Multidisciplinary Approaches for Hepatocellular Carcinoma Early Diagnosis,” will be presented Wednesday at 3 p.m. as part of Scientific Session 34, IO Frontiers 2.
HCC is a leading cause of cancer-related deaths around the world, and there has been a significant increase in HCC-related deaths in the United States in the past 40 years. Cirrhosis is the key risk factor for HCC; among HCC patients, 90% have underlying cirrhosis.
Cirrhosis is typically monitored with regular ultrasounds and blood tests that track alpha fetoprotein (AFP) levels, according to presenter Hongyu Wang, MD, PhD, assistant professor in the department of diagnostic and interventional imaging at the University of Texas Health Science Center at Houston. High AFP levels could indicate HCC, but the results are not sensitive enough or reliable enough, Dr. Wang said.
Dr. Wang’s team wanted to improve HCC early diagnosis by using AI to collect and analyze a variety of information from patients. Oftentimes, cirrhosis patients have an inconsistent type and timeline of images, she said.
“When the patients come in, maybe this time they’re prescribed a CT scan and the next time they may change to an MRI. …. Quite a lot of patient data are not consistent with the same kind of scanning method,” Dr. Wang explained.
From their affiliated hospital’s database, the team collected data on 220 patients that had been diagnosed with HCC with underlying cirrhosis. They then used the cycle-GAN AI model to transform images between MRI and CT and used a hybrid three-dimensional deep convolutional neural network (3DCNN) to extract image features and classify the HCC. The study showed that the cycle-GAN model could learn most CT and MR image styles during image transformation. The accuracy of HCC prediction was 0.8625, and the precision for HCC prediction was 0.58.
They also studied the differentially expressed mRNA, mRNA-encoded protein and overall survival using the data of 364 HCC patients from The Cancer Genome Atlas. They chose six best-performing genes as their targets, confirming four of the genes as prognostic markers for liver cancer survival. They were unfavorable ANXA2, FUBP1, CDK1 and favorable KNG1.
“I think this work is really significant if we can achieve this,” Dr. Wang said. “It will bring a very significant clinical application, and we can improve the current method and better classify the patients into high risk or low risk.”
Eventually, she said she hopes they can “build up a system so users can plug in data and get this information that has integrated all these different factors of the image data and the protein mRNA biomarkers … so everybody can use this system to enhance your work, increase your accuracy, save your time and energy, and thus improve the diagnosis.”
This model could also expand to detect other cancers, she said.
AI is the future of data analysis, Dr. Wang said, providing more accurate detail and analysis to “help physicians and scientists to gather clues.”
“AI has dramatically changed the current medical services, especially for radiologists. The workload for radiologists has dramatically increased, too. Therefore, AI has the capability to increase the accuracy, reduce error, and improve the performance of the diagnosis and treatment.”
