Rolling in the Deep (Learning)
A new machine learning model can support the differential diagnosis of tricky liver lesions
George Lee | | Quick Read
Artificial intelligence (AI)’s role in pathology is rapidly picking up pace – particularly as a decision-making aid in tricky diagnostic situations. Hepatocellular nodular lesions (HNLs) are one such situation; differentiating between HNL types, particularly high-grade dysplastic nodules and hepatocellular carcinoma, has historically proved challenging. Can AI help?
To find out, Na Cheng and colleagues evaluated four different neural networks’ ability to analyze a set of 1,115 whole-slide HNL images samples from 738 patients (1). Each specimen was also reviewed by two to three subspecialists. The performance of the optimal model, known as the hepatocellular-nodular artificial intelligence model (HnAIM), was compared with diagnoses made by nine pathologists to examine its predictive efficiency.
How did it do? For biopsy specimens, HnAIM had a higher agreement rate with subspecialists’ majority opinion than the pathologists – a promising result for a potential diagnostic assistance tool. Furthermore, its strong patch-level recognition means that tools such as HnAIM may offer useful diagnostic support in cases of fragmented, low-volume, or low-quality biopsies.
- N Cheng et al., Gastroenterology, [Online ahead of print] (2022). PMID: 35202643.