‘Gist’ processing of images enhances AI’s breast cancer risk assessment
Machine learning—combined with radiologists’ intuitive, or “gist” processing—is more accurate in assessing breast cancer risk than either approach alone.
Deep learning models, such as convolutional neural networks (CNNs), enable automatic screening of life-threatening breast cancers at an earlier, more curable stage. However, CNNs rely on large annotated datasets for clinical diagnosis as input.
Gist is the memory representation of the bottom-line meaning of an experience. Gist-based intuition is a human visual ability, an advanced form of reasoning that is mainly unconscious and develops with experience. However, it has not been used in neural networks for screening mammograms even though it enables the human visual system to readily extract meaningful information. For instance, radiology experts can classify mammograms as normal or abnormal at above-chance levels after less than one second exposure to the image.
Researchers, from IBM Watson Health, Ohio State University and elsewhere, hypothesized that a CNN plus the gist information from radiology experiences would provide more accurate results than either method on its own. Their goal was not to create a novel CNN for screening but use machine learning models that were pretrained and add the gist response in input to models already in use.
They used a mammogram input data set containing 220 digital mammograms from 110 unique patients, and a transfer learning approach to combine the gist information with features from a CNN classifier.
The combined method was statistically more accurate than either the radiologist experts or the CNN alone, including reducing false positive rates and improving the true positive rate.
“Our approach incorporates the informed decisions of radiologists’ who have years of education and experience with the image-analysis and pattern-recognition capabilities provided by deep CNNs and machine learning. The combination of the parts is better than either solution alone in our training data, as each input captures different features,” the researchers stated.
The study was published in arXiv.org, part of the Cornell University Library.
The researchers also suggested that this method can be applied more universally.
“Similar solutions utilizing both inputs may prove useful in other problem domains, especially in the medical field where trained professionals often work with computer-aided detection systems and may pick up on different signals than does a deep neural network. In these cases, we suggest human and CNN collaborative problem solving,” they concluded.