LLNL Uses ‘Self-Trained’ Deep Learning for Better Diagnosis
Improving disease diagnosis is one of the most popular applications for deep learning researchers – and now, they’re looking to streamline the process: new research from Lawrence Livermore National Laboratory (LLNL) uses new “self-training” deep learning models to diagnose diseases from X-ray imaging.
Data acquisition and labeling are big pain points for AI applications in diagnosis: health data is hard to acquire, costly when available and difficult to label. This, in turn, impedes the AI models’ ability to diagnose, as they lack the data necessary to make inferences.
“Building predictive models rapidly is becoming more important in health care,” said Jay Thiagarajan, a computer scientist at LLNL. “There is a fundamental problem we’re trying to address. Data comes from different hospitals and it’s difficult to label — experts don’t have the time to collect and annotate it all. It’s often posed as a multi-label classification problem, where we are looking at the presence of multiple diseases in one shot. We can’t wait to have enough data for every combination of disease conditions, so we built a new technique that tries to compensate for this lack of data using regularization strategies that can make deep learning models much more efficient, even with limited data.”
The new framework addresses those pain points. In essence, a “teacher” model is trained on a limited amount of labeled data; then, the “teacher” model trains a “student” model using labeled and unlabeled data. Through this process – combined with data augmentation and confidence tempering – the resultant “student” model outperforms its instructor; in fact, the researchers claim that the resulting models can perform as well as – or better than – models trained on much larger labeled datasets.
The test case: chest X-ray datasets used to diagnose a range of heart conditions, such as edema and cardiomegaly. Using 85% less labeled data, the researchers achieved the same performance as other state-of-the-art models.
“When you have limited data, improving the capability of models to handle data it hasn’t seen before is the key aspect we have to consider when solving limited data problems,” he explained. “It’s not about picking Model X versus Model Y, it’s about fundamentally changing the way we train these models, and there’s a lot more work that needs to be done in this space for us to achieve meaningful diagnosis models for real-world use cases in healthcare.”
The model isn’t perfect: X-rays are a straightforward test case, but other diagnoses may not be as simple. The model is also prone to overconfidence, hence the confidence tempering measures in its design. As the team moves forward, it will be integrating domain knowledge to make the model more broadly capable.