Computer Predicts Cancer Survival Better than Physicians

When pathologists peek through microscopes to see if a person has signs of cancer, even the best experts often have trouble agreeing about what they see. That 80-year-old routine could finally get a boost from a new computer program that has proved better than humans at predicting survival among breast cancer patients.

The "C-path" program didn't just screen microscopic samples based on cell features that human pathologists consider important. Instead, it came up with more than 6,000 characteristics of cancer on its own, and used those criteria to predict the severity of breast cancer among two groups of women. Its success could someday lead to programs capable of identifying the best drugs or treatments for certain patients.

"I think it would be interesting to see if you could really decide which patients are the best candidates for a treatment, based on learning a new model relevant to a specific treatment," said Andrew Beck, a pathologist at Harvard University. "Then you could impact patient care."

A learning computer

Beck formerly represented one of several Stanford University physicians who worked with computer scientists to create C-path over the past three years. They trained the program by using data from former cancer patients in the Netherlands, so that C-path could identify a new set of cell features related to better or worse cancer survival.

"A lot of the work of looking at a [microscope] slide comes from knowing what to ignore, and knowing to focus on features that are key," Beck told InnovationNewsDaily. "A computer isn't trained to do that, and so it has an open mind to finding associations that may not have been identified before."

But C-path still suffered from the same problem that plagues all of today's existing computer-driven programs or robots — it had trouble telling certain visual patterns apart. By contrast, the human eye and brain can learn to identify different cell types related to cancer.

For Subscribers

FOR SUBSCRIBERSAbdul El-Sayed tries to woo MAHA voters as Democrats seek inroads with RFK’s supporters

FOR SUBSCRIBERSKratom-like substances are popping up in a typically benign supplement

FOR SUBSCRIBERSGOP super PAC increases its spending for Sen. Jon Husted in tough Ohio race

FOR SUBSCRIBERSRecords reveal tumultuous marriage between slain New York Times executive and wife

FOR SUBSCRIBERSFrom clean energy warrior to ‘wall of shame,’ James Talarico reverses on a fossil fuel money pledge

FOR SUBSCRIBERSCongressional staff planned official trip with luxury stops in Italy and France weeks before election

FOR SUBSCRIBERSA harrowing ordeal aboard an Israel-bound plane

FOR SUBSCRIBERSOpenAI CEO Sam Altman to skip congressional hearing on rogue AI agents

FOR SUBSCRIBERSKey DOJ lawyers departed Trump ‘grand conspiracy’ probe amid pressure to produce, source says

FOR SUBSCRIBERSTrump’s new AI chatbot often links to a site launched under Clinton 26 years ago

FOR SUBSCRIBERSAt MAHA summit, industry CEOs the movement once vilified take the stage

FOR SUBSCRIBERSThis Paralympian has two goals: Win medals and give away $500,000 to her teammates

FOR SUBSCRIBERSCornell rape allegation spotlights stringent New York state laws that have long drawn criticism

FOR SUBSCRIBERSThe debate over AI’s future rages on both coasts

FOR SUBSCRIBERSDecades-old march dispute reignites in Northern Ireland, threatening political crisis

Distinguishing between different cells matters, because runaway cancer mutations typically take place in epithelial tissue cells lining the outside and inside of human body structures. Connective tissue, such as blood vessels, plays a different role by influencing cancer growth and response to treatment.

Putting C-path to the test

To help C-path, the Stanford team hand-marked 158 samples to identify epithelial tissue versus connective tissue. After the program trained on 248 breast cancer cases from the Netherlands Cancer Institute, it took on a real test by looking at 286 past cases from Vancouver General Hospital in Canada.

Pathologists usually sort patients into groups with a high or low risk of death. If many more patients in the high-risk group end up dying compared to the low-risk group – measured as the "hazard ratio" – that means the predictions had proven accurate.

For Subscribers

FOR SUBSCRIBERSAbdul El-Sayed tries to woo MAHA voters as Democrats seek inroads with RFK’s supporters

FOR SUBSCRIBERSKratom-like substances are popping up in a typically benign supplement

FOR SUBSCRIBERSGOP super PAC increases its spending for Sen. Jon Husted in tough Ohio race

FOR SUBSCRIBERSRecords reveal tumultuous marriage between slain New York Times executive and wife

FOR SUBSCRIBERSFrom clean energy warrior to ‘wall of shame,’ James Talarico reverses on a fossil fuel money pledge

FOR SUBSCRIBERSCongressional staff planned official trip with luxury stops in Italy and France weeks before election

FOR SUBSCRIBERSA harrowing ordeal aboard an Israel-bound plane

FOR SUBSCRIBERSOpenAI CEO Sam Altman to skip congressional hearing on rogue AI agents

FOR SUBSCRIBERSKey DOJ lawyers departed Trump ‘grand conspiracy’ probe amid pressure to produce, source says

FOR SUBSCRIBERSTrump’s new AI chatbot often links to a site launched under Clinton 26 years ago

FOR SUBSCRIBERSAt MAHA summit, industry CEOs the movement once vilified take the stage

FOR SUBSCRIBERSThis Paralympian has two goals: Win medals and give away $500,000 to her teammates

FOR SUBSCRIBERSCornell rape allegation spotlights stringent New York state laws that have long drawn criticism

FOR SUBSCRIBERSThe debate over AI’s future rages on both coasts

FOR SUBSCRIBERSDecades-old march dispute reignites in Northern Ireland, threatening political crisis

C-path's predictions led to a high-risk group with 75 percent increased annual risk of death compared to the low-risk group. That handily beat the usual pathologist grading, in which the high-risk group had a 20 percent increased annual risk of death compared to the low-risk group.

Still, the program first had to train on a small, separate set of samples from Vancouver General Hospital, according to a paper published by Beck and colleagues in the Nov. 9 issue of the journal Science Translational Medicine.

The future of medicine

Such computer programs won't put pathologists out of their jobs anytime soon. Instead, they may prove useful tools that raise the accuracy of cancer prognosis, said David Rimm, a pathologist at Yale University. Rimm wrote a related Science Translational Medicine paper about C-path as an independent expert.

"Now, you can envision a time when you put the slide under the microscope and a robot does everything a pathologist does, gives all info that pathologist usually gives, and replaces the pathologist," Rimm said. "They're not anywhere close to that yet."

Beck and his Stanford colleagues have begun working on scaling up C-path to look at entire microscope slides rather than a small sample — a task that requires much more computational power. They also want to figure out if the computer approach requires different medical institutions to standardize how they create microscope slides.

Having computer programs analyze cells for signs of cancer could also complement the genomics approach that tries to identify DNA features related to cancer. In cases where genetics play a big role, the computer-driven approach could tie the genetics together with the changes in cell appearances.

"Instead of quantifying features by eye, you can quantify a broad spectrum of features and find which are associated with important molecular changes," Beck said.

You can follow InnovationNewsDaily Senior Writer Jeremy Hsu on Twitter @ ScienceHsu. Follow InnovationNewsDaily on Twitter @ News_Innovation, or on Facebook.

• 10 Sci-Fi Predictions That Came True
  
• INFOGRAPHIC: Post 9/11 Medical Advances
  
• 9 Cyborg Enhancements Available Right Now