Machine learning holds its own in breast cancer

2019 07 27 00 32 8796 Breast Cancer Cell2 400

A machine-learning method was able to predict biomarker expression in breast cancer through analyses of hematoxylin-eosin (H&E) stained images alone, on par with standard immunohistochemistry, in a single-center, retrospective study of publicly available cases, Israeli researchers reported on July 26 in JAMA Network Open.

The group evaluated the method -- called morphological-based molecular profiling (MBMP) -- for predicting 19 biomarkers based on 20,600 digitized, publicly available H&E images taken of 5,356 breast cancer patients seen at the Vancouver General Hospital in British Columbia. The method matched immunohistochemistry (IHC) staining for predicting biomarker expression for at least half of the patients in the study, which was funded by grants from Israeli governmental bodies and foundations.

For example, when it came to predicting estrogen-receptor expression (the status of which is needed to guide therapy), the method demonstrated accuracy of 91% to 92%, with results in line with what has been reported for traditional IHC, according to the researchers (JAMA Netw Open, July 2019, Vol. 2:7, e197700).

Prediction of estrogen-receptor expression
  H&E images Traditional IHC, ranges reported in other trials
Study cohort 1 Study cohort 2
Positive predictive value 97% 98% 91%-98%
Negative predictive value 68% 76% 51%-78%
Accuracy 91% 92% 81%-90%

Furthermore, the researchers believe they can improve upon the accuracy of the method. However, it's early days yet for machine-learning applications.

"Image analysis by machine learning is gaining ground for various applications in pathology but has not been proposed to replace chemical-based assays for molecular detection," acknowledged Gil Shamai, from the department of electrical engineering at Technion -- Israel Institute of Technology in Haifa, and colleagues.

Could that change in the future? The researchers see a window of opportunity in light of the downsides of immunohistochemistry using monoclonal antibodies -- which they described as the "workhorse of molecular genotyping" -- such as high costs, the subjective nature of pathology interpretations, and dependence on tissue handling protocols and expert lab technicians.

"Morphological-based molecular profiling could be used as a general approach for mass-scale molecular profiling based on digitized H&E-stained images, allowing quick, accurate, and inexpensive methods for simultaneous profiling of multiple biomarkers in cancer tissues," they wrote.

Page 1 of 12
Next Page