Case study: AI models aiding chronic cholestasis detection

A cholestatic condition is characterized by decreased bile flow as substances normally excreted into bile are retained. This is often due to impaired hepatocytes or obstruction of bile ducts caused by primary damage to the biliary epithelium. The majority of adult patients with chronic cholestasis have primary sclerosing cholangitis (PSC), a long-term progressive liver disease that affects both small and large bile ducts.

PSC is one of the leading causes of liver transplantation in the Nordic countries. It damages the bile ducts through inflammation and ultimately causes cirrhosis, leading to loss of liver function as the disease progresses. There are currently no effective treatments, as the disease can even reoccur after liver transplantation, with some patients having to undergo several transplantations.

Traditional methods for K7 detection and analysis

Cytokeratin 7 (K7) is an important marker for chronic cholestasis prognosis. In the normal liver, K7 expression occurs in the biliary epithelium, whereas hepatocytes remain negative. However, in chronic cholestasis, periportal hepatocytes and intermediate hepatobiliary cells stain positive for K7. Hence, the expression of K7 in the liver is commonly known as an indicator of bile duct injury.

However, traditional methods for K7 detection and analysis are subjective and can lead to significant result variability. Dr. Nelli Sjöblom, MD, Pathologist at Helsinki University Hospital, aims to solve this challenge using AI models to locate and quantify K7-positive hepatocytes for accurate and reliable results. We interviewed Nelli to learn more about her new publication on the topic and her experience using Aiforia^® Create, Aiforia's cloud-based AI development tool.

Nelli has developed multiple projects regarding PSC with Aiforia; read more and view a short video showcasing her work below. 

Tell us a bit about the project or research work you used Aiforia’s software for

Nelli: "We aimed to develop an automated image analysis tool to assess the amount of K7-positive hepatocytes in any liver biopsy specimen. The quantity of K7-positive cells in a liver biopsy specimen reflects the amount of chronic cholestasis and is also a predictive marker in some liver diseases. Cholestasis can be caused by multiple different liver diseases and liver failure. The distinction between K7-positive biliary epithelium and the cholestatic hepatocytes was not an easy task to be automatized in the liver with some of the conventional machine learning tools because both of them are stained very similarly in the K7 staining."

What do you think is most unique about the project behind this publication?

"To our knowledge, the amount of chronic cholestasis and its predictive value has not been studied before in a cohort of patients with primary sclerosing cholangitis (PSC) that was chosen as a ‘validation’ cohort for our AI model. Our team has access to a unique PSC registry (Helsinki University Hospital), including all the patient’s clinical data and their histological liver specimens, which gave us the opportunity to investigate and utilize their samples in the development of this AI model. More prognostic markers are needed for this rare disease, for which liver transplantation surgery is the only effective treatment option. In addition, digital pathology is taking over our traditional diagnostic methods (e.g., the microscope), so we wanted to see if developing tools for diagnostic assistance would also work in analyzing these specifically stained liver specimens."

How was your experience learning to use Aiforia's software? Have you worked with AI previously?

"I had not worked with AI or Aiforia's software before, but I had read about automated image analysis and methods developed to assist pathologists working in fully digitalized pathology laboratories. Thus, I immediately became interested when my supervisors Prof. Johanna Arola and Prof. Martti Färkkilä offered me this project as part of my thesis. 

The experience was great – the platform was rather intuitive and I find it very important that the people (the pathologists) using these tools will be involved in the development of such tools as well."

"The automation makes the image analysis and histological interpretation fast and objective. The aim was to show that the methodology described in the article is non-inferior to a human pathologist, and that we can rely on the results produced by the AI model. This is valuable information regarding the future use of such tools."

Why do you think AI-based image analysis tools have not previously been used in the assessment of chronic cholestasis? Do you find it is worth investing time into?

"Not many commercial AI models have been developed for purposes other than to aid in diagnostics of the most common malignancies, such as breast cancer or prostate cancer. But I am sure when they become more accessible and affordable, more pathologists will be able to build AI models specifically designed for their needs.

Objective and rapid tools are needed in different fields of pathology, especially when it comes to research and drug development (treatment response assessment). I believe that is where the AI models truly prove their worth."

What are your next steps after this publication; do you plan to use AI for further research or other projects?

"The next phase is to apply this tool to the previously described PSC cohort and see whether the results correlate with disease progression and whether chronic cholestasis will indicate a poorer prognosis. Development of other models for different staining methods of the liver is ongoing."

[Video] Deep neural networks in diagnosis and disease progression of PSC

Watch a video showcasing multiple projects over the course of Nelli’s work on Primary Sclerosing Cholangitis.

1. 0:09 Implementing machine learning and supervised learning in labeling the training set of images
2. 1:03 New classification of histology based on an algorithm in PSC and its impact on prognosis
3. 2:00 Indicators of chronic cholestasis (CK-7) in the assessment of PSC based on deep learning networks
   
   

Read more about the beginning of Nelli’s work here: Case study: evaluating prognostic indicators in primary sclerosing cholangitis with AI

References

Milkiewicz, M. et al. (2016, December 23). Impaired hepatic adaptation to chronic cholestasis induced by primary sclerosing cholangitis. Scientific Reports, 6. https://doi.org/10.1038/srep39573 

Poupon, R. et al. (2000). Chronic cholestatic diseases. J. Hepatol, 32(1), 129-140. https://doi.org/10.1016/s0168-8278(00)80421-3