A self-taught artificial intelligence system can identify a common form of lung cancer from microscopy images and determine the likelihood that it will return after treatment as well as overall patient survival.
The system has potential to one day relieve pathologists from the time-consuming manual interpretation of stained tissue sections.
It also holds considerable prognostic and biomarker potential, and could be used to predict drug responses if trained with a larger dataset.
“Lung tissue samples can now be analyzed in minutes by our computer program to provide fairly accurate predictions of whether their cancer will return, predictions that are better than current standards of care for making a prognosis in lung adenocarcinoma,” said study co-senior investigator Aristotelis Tsirigos, PhD, from the NYU Grossman School of Medicine.
The researchers report their findings in the journal Nature Communications.
Tsirigos and co-workers initially used their histomorphological phenotype learning (HPL) system to analyze 541 slides of lung adenocarcinoma from 452 patients who were on the Cancer Genome Atlas.
This type of lung cancer has many subtypes and different kinds of features, with tumor structure and form highly predictive of patient outcomes.
Images of the slides were digitally scanned and broken into 432,231 small quadrants or tiles that the system grouped into distinct tissue patterns, effectively cataloguing the highly diverse morphologies of the tumor type.
The team identified 46 distinct characteristics from normal and diseased tissue that they labelled histomorphological phenotype clusters, a subgroup of which were significantly associated with either recurrent cancer or long-term survival.
The team then validated their findings in an analysis of tissue images from 276 men and women treated for adenocarcinoma at NYU Langone between 2006 and 2021.
The HPL could accurately differentiate between similar lung cancers, adenocarcinoma and squamous cell cancers, 99% of the time.
It was 72% accurate at predicting the likelihood and timing that cancer would return after treatment, beating the 64% accuracy for predictions made by pathologists directly examining the same patients’ tumor images.
“Our new histomorphological phenotype learning program has the potential to offer cancer specialists and their patients a quick and unbiased diagnostic tool for lung adenocarcinoma that, once further testing is complete, can also be used to help validate and even guide their treatment decisions,” said researcher Nicolas Coudray, PhD, a bioinformatics programmer at the NYU Grossman School of Medicine.
He added: “Patients, physicians, and researchers know they can rely on this predictive modeling because it is self-taught, provides explainable decisions, and is based only on the knowledge drawn specifically from each patient’s tissue, including such features as its proportion of dying cells, tumor-fighting immune cells, and how densely packed the tumor cells are, among other features.”
The team believes the algorithm will become increasingly accurate as more data is added, due to its self-learning nature. They hope one day it could be used to give patients a score between zero and one that reflects their chances of survival and tumor recurrence for up to five years.
The programming code has been freely posted online and the researchers hope to make their HPL tool also available after further testing.
