NAFLD case study: assessing liver histology with AI

This case study describes the use of AI in studying nonalcoholic fatty liver disease (NAFLD) and its capability to segment structures in liver histology.
Written by Aiforia

What is your job and what does your research focus on?

Interview with Sami QadriSami: I am an MD PhD student at the University of Helsinki and I work in professor Hannele Yki-Järvinen’s research group. Our main focus of investigation currently is nonalcoholic fatty liver disease (NAFLD), which has emerged as the most common liver disease  worldwide. 

My personal studies aim to shine light on the effects of particular genetic determinants on NAFLD and the mechanisms by which NAFLD arises in these genetically predisposed subjects. A part of this work entails harnessing AI to analyse liver biopsies in completely new ways worldwide. 

What does Aiforia enable you to do that you otherwise could not (or would find difficult) to with traditional or other methods?

Most importantly, Aiforia enables segmentation of both normal and pathological structures in liver histology, in a way that allows us to accurately quantitate these features in multiple different ways. The traditional method of analysing liver histology, i.e. visual assessment by pathologists, is rather problematic from the standpoint of medical research. 

First, pathologists can at best give us semi-quantitative assessments with regard to the amount or extent of pathology that is present. Aiforia grants us access to truly quantitative measures with continuous metrics instead of arbitrary grading. 

Second, we know that the ‘internal calibration’ among pathologists may differ, leading to marked observer-related variability in assessments. For research purposes these inconsistencies are naturally problematic. Lastly, with Aiforia we will gain access to completely new kinds of metrics, like quantifying the shape or size of individual lesions, or their spatial relationships.

What were your expectations of using deep learning AI?

It was difficult to weigh my expectations as I had never worked with AI before. After seeing the platform’s capabilities, however, I was thoroughly excited from the onset. I am glad to say that all of those expectations have been met and even exceeded!

Did you know anything about AI before starting your work with Aiforia?

I knew about the basics of AI in general, but very little in practise. One misconception I had was the thought that it would take lots of training material for an AI model to be effective. This turned out to be incorrect, as Aiforia’s image segmentation becomes surprisingly specific with just a few annotated examples. Of course for a robust AI model you will need variability, but a rough working model can be devised in a matter of minutes.