Using smart software to analyse lung cancer tissue samples makes the diagnostic process more accurate, a team of researchers at Stanford University School of Medicine has found.

The team discovered that machine-learning makes it easier to identify and differentiate between two different types of lung cancer by looking at the essential features of each disease. These smart machines were also able to predict survival times more accurately than human pathologists, as well as to assess the tumours by stage and grade.

A difference of opinion

Current pathology is limited by having to use the human eye, as well as the pathologist’s own experience and judgement. Even highly experienced and skilled pathologists will disagree with each other around 40% of the time, which means patients don’t always get the right treatment in the right time frames.

Increasingly, path labs will rely on image analysis software developed by companies like Bitplane to pin down diagnoses and improve patient care and outcomes.

This Stanford study looked only at two types of lung cancer – adenocarcinoma and squamous cell carcinoma – but the technology can be applied to many other forms of cancer. It’s believed that the learning software can examine thousands of disease-specific features, rather than the several hundred that pathologists look for.

Grading the disease

Pathologists have, for decades now, graded the severity of a cancer by looking at the tissue through a light microscope. The severity of the disease was assessed by looking at how much abnormal tissue there was, as well as the shapes and sizes of individual cells. They also look at whether the cancer has spread.

These assessments are used to decide the treatment path, as well as to predict the outcome. In the cases of adenocarcinoma and squamous cell carcinoma, it’s sometimes difficult to tell the two subtypes apart.

It’s also difficult to make accurate predictions about outcome – half of stage-1 adenocarcinoma patients die within five years of diagnosis, whereas 15% live for ten years or more – as outcomes can vary widely.

More insight into the future

It seems that digital image analysis can offer a better prediction as well as a better diagnosis. The Stanford team used almost 2,200 images of adenocarcinoma and squamous cell carcinoma from The Cancer Genome Atlas database to train their computers.

This database also contains the eventual outcomes for these patients, as well as the grade and stage of the cancers.

By training their machines to look at almost 10,000 cancer-specific characteristics, rather than the mere hundreds that humans look at, the team found that the computers could perceive and assess miniscule differences between samples.

This meant that the computers could rapidly identify the subtype of cancer and predict how long each patient would live after diagnosis. The software could accurately differentiate between the long-term and short-term survivors whose details were on another database.

Better treatments

As this emerging technology can identify and assess previously unseen characteristics to diagnose cancers and predict outcomes, it’s likely to help researchers to understand the molecular processes involved in the disease.

Even more exciting is the possibility of linking this tech to the fields of genomics and proteomics as genetic mutations and protein expressions are central to understanding cancer.