Deep learning models superior to machine learning at brain imaging analysis
Deep learning models can extract more insights to create better representations of the brain than standard machine learning models.
That’s what researchers at Georgia State University say and chalk up to its superior ability to determine patterns and distinguish features from large amounts of data within complex brain imaging, and to continue learning on their own from data they analyze.
“Deep learning is more likely to replace many other methods in recognition tasks from large inputs, such as tomographies and other radiological measurements. Especially in the areas where a large quantity of historical data is already available but compact and precise biomarkers are not fully understood and we have to rely on highly trained human experts and their judgement,” lead author Anees Abrol, research scientist at the Center for Translational Research in Neuroimaging and Data Science (TReNDS) and co-author Sergey Plis, associate professor of computer science and director of machine learning at TReNDS at Georgia State University, told HCB News.
These findings contrast with skeptics who have found little difference between deep learning and standard machine learning when assessing brain images. Abrol says such comparisons are marred by their use of pre-processed data, as this eliminates the main feature of deep learning, its ability to learn from data with little to no preprocessing.
For their study, he and his colleagues applied representative models of both technologies to sample sizes from 100 to 10,000. They found that if trained properly, all deep-learning approaches have the potential to generate superior representations of the human brain. The reason for this can be explained through reverse analysis, which shows that the models continue to learn how to identify significant brain biomarkers from the data they process.
The authors say that deep learning will not fully replace all possible tasks performed with standard machine learning. While acknowledging that training deep learning models require large amounts of information, they also assert that such training enables the technology to assess reams of complex data just as effectively as when it answers simple questions.
“Every methodology has its niche and DL models will likely dominate one where data dimensionality is large and difficult to comprehend by a human expert, such as imaging in radiology, big data electronic health records, and genomic data,” said Abrol and Plis.
They add that furthering research in this area requires investments in the deep learning field and biomedical data; establishing forums and venues where biomedical experts can discuss problems they believe deep learning can solve using annotated datasets; and for each important problem establish as large as possible benchmark datasets with a clear metric of success.
The findings were published in Nature Communications.