From the Journals

Deep learning system outmatches pathologists in diagnosing liver lesions

View on the News

Work smarter not harder

As the prevalence of hepatocellular carcinoma (HCC) continues to rise, the early and accurate detection and diagnosis of HCC remains paramount to improving patient outcomes. In cases of typical or advanced HCC, an accurate diagnosis is made using CT or MR imaging. However, hepatocellular nodular lesions (HNLs) with atypical or inconclusive radiographic appearances are often biopsied to achieve a histopathologic diagnosis. In addition, accurate diagnosis of an HNL following liver resection or transplantation is important to long-term surveillance and management. An accurate histopathologic diagnosis relies on the availability of experienced subspecialty pathologists and remains a costly and labor-intensive process that can lead to delays in diagnosis and care.

Hannah P. Kim, MD, MSCR, is an assistant professor in the division of gastroenterology, hepatology, and nutrition in the department of medicine at Vanderbilt University Medical Center, Nashville, Tenn.

Dr. Hannah P. Kim

In this study, Cheng et al. developed a deep learning system to differentiate histopathologic diagnoses of various HNLs, normal liver, and cirrhosis. Their model, hepatocellular-nodular artificial intelligence model (HnAIM), accurately classified various liver histology slides with an AUC of 93.5% using an external validation cohort. When compared to even the most experienced subspecialty pathologists, HnAIM demonstrated superior HNL histopathologic diagnostic accuracy. Utilization of HnAIM to either make or aid in the diagnosis of HNLs can lead to more accurate diagnoses in a more efficient and timely manner and has the potential to provide subspecialty care in areas that lack subspecialty pathologists. If this model is further validated, HnAIM may be used to improve the quality of care we are able to provide to our patients, ultimately with the ability to improve our diagnosis of HCC, prevent delays in treatment, and improve patient outcomes.

Hannah P. Kim, MD, MSCR, is an assistant professor in the division of gastroenterology, hepatology, and nutrition in the department of medicine at Vanderbilt University Medical Center, Nashville, Tenn. She has no conflicts of interest.



A new deep learning system can classify hepatocellular nodular lesions (HNLs) via whole-slide images, improving risk stratification of patients and diagnostic rate of hepatocellular carcinoma (HCC), according to investigators.

While the model requires further validation, it could eventually be used to optimize accuracy and efficiency of histologic diagnoses, potentially decreasing reliance on pathologists, particularly in areas with limited access to subspecialists.

In an article published in Gastroenterology, Na Cheng, MD, of Sun Yat-sen University, Guangzhou, China, and colleagues wrote that the “diagnostic process [for HNLs] is laborious, time-consuming, and subject to the experience of the pathologists, often with significant interobserver and intraobserver variability. ... Therefore, [an] automated analysis system is highly demanded in the pathology field, which could considerably ease the workload, speed up the diagnosis, and facilitate the in-time treatment.”

To this end, Dr. Cheng and colleagues developed the hepatocellular-nodular artificial intelligence model (HnAIM) that can scan whole-image slides to identify seven types of tissue: well-differentiated HCC, high-grade dysplastic nodules, low-grade dysplastic nodules, hepatocellular adenoma, focal nodular hyperplasia, and background tissue.

Developing and testing HnAIM was a multistep process that began with three subspecialist pathologists, who independently reviewed and classified liver slides from surgical resection. Unanimous agreement was achieved in 649 slides from 462 patients. These slides were then scanned to create whole-slide images, which were divided into sets for training (70%), validation (15%), and internal testing (15%). Accuracy, measured by area under the curve (AUC), was over 99.9% for the internal testing set. The accuracy of HnAIM was independently, externally validated.

First, HnAIM evaluated liver biopsy slides from 30 patients at one center. Results were compared with diagnoses made by nine pathologists classified as either senior, intermediate, or junior. While HnAIM correctly diagnosed 100% of the cases, senior pathologists correctly diagnosed 94.4% of the cases, followed in accuracy by intermediate (86.7%) and junior (73.3%) pathologists.

The researchers noted that the “rate of agreement with subspecialists was higher for HnAIM than for all 9 pathologists at distinguishing 7 liver tissues, with important diagnostic implications for fragmentary or scarce biopsy specimens.”

Next, HnAIM evaluated 234 samples from three hospitals. Accuracy was slightly lower, with an AUC of 93.5%. The researchers highlighted how HnAIM consistently differentiated precancerous lesions and well-defined HCC from benign lesions and background tissues.

A final experiment showed how HnAIM reacted to the most challenging cases. The investigators selected 12 cases without definitive diagnoses and found that, similar to the findings of three subspecialist pathologists, HnAIM did not reach a single diagnostic conclusion.

The researchers reported that “This may be due to a number of potential reasons, such as inherent uncertainty in the 2-dimensional interpretation of a 3-dimensional specimen, the limited number of tissue samples, and cognitive factors such as anchoring.”

However, HnAIM contributed to the diagnostic process by generating multiple diagnostic possibilities with weighted likelihood. After reviewing these results, the expert pathologists reached consensus in 5 out of 12 cases. Moreover, two out of three expert pathologists agreed on all 12 cases, improving agreement rate from 25% to 100%.

The researchers concluded that the model holds the promise to facilitate human HNL diagnoses and improve efficiency and quality. It can also reduce the workload of pathologists, especially where subspecialists are unavailable.

The study was supported by the National Natural Science Foundation of China, the Guangdong Basic and Applied Basic Research Foundation, the Natural Science Foundation of Guangdong Province, and others. The investigators reported no conflicts of interest.

Next Article: