Deep learning shows a lot promise in health care, especially in medical imaging, where it can be used to improve the speed and accuracy of diagnosing the patient’s condition. But it also faces a serious hurdle: the shortage of labeled training data.
To overcome this hurdle, scientists have explored various solutions with varying degrees of success. In a new document, artificial intelligence Google researchers suggest a new technique it uses self-supervised learning to train deep learning models for medical imaging. Early results show that the technique can reduce the need for annotated data and improve the performance of deep learning models in medical applications.
Convolutional neural networks have proven to be very efficient in computer vision tasks. Google is one of several organizations that has explored its use in medical imaging. In recent years, the company’s research arm has built several medical imaging models in domains such as ophthalmology, dermatology, mammography, and pathology.
“There is a lot of excitement in applying deep learning to health, but it remains a challenge because highly accurate and robust DL models are needed in an area like healthcare,” said Shekoofeh Azizi, AI resident at Google search and lead author of the self-supervised paper.
One of the key challenges of deep learning is the need for huge amounts of annotated data. Great neural networks require millions of labeled examples to achieve optimum accuracy. In medical settings, data labeling is a complicated and costly undertaking.
“Acquiring these ‘labels’ in the medical field is challenging for a variety of reasons: it can be time-consuming and costly for clinical experts, and data must meet relevant privacy requirements before being shared,” said Azizi.
For some conditions, examples are scarce, to begin with, and in others, such as breast cancer screening, it may take many years for clinical results to manifest after a medical image is taken.
Complicating the data requirements of medical imaging applications further are changes in distribution between training data and delivery environments, such as changes in patient population, disease prevalence or presentation, and the medical technology used to image capture, added Azizi.
A popular way to address medical data shortages is to use supervised pre-training. In this approach, a convolutional neural network is initially trained on a tagged image dataset, such as ImageNet. This step tunes the model layer parameters to the general models found in all types of images. The trained deep learning model can then be fine-tuned to a limited set of examples labeled for the target activity.
Several studies have shown supervised pre-training to be useful in applications such as medical imaging, where tagged data is scarce. However, even supervised pre-workout has its limitations.
“The common paradigm for training medical imaging models is learning transfer, in which models are first pre-trained using supervised learning on ImageNet. However, there is a large dominance shift between natural images in ImageNet and medical images, and previous research has shown that such supervised pre-training on ImageNet may not be optimal for developing medical imaging models, “said Azizi. .
Self-paced learning has emerged as a promising research area in recent years. In self-supervised learning, deep learning models learn representations of training data without the need for labels. When done correctly, self-supervised learning can be of great benefit in areas where tagged data is scarce and untagged data is abundant.
Outside of medical circles, Google has developed several self-supervised learning techniques to train neural networks for computer vision tasks. Among these is the Simple Framework for Contrastive Learning (SimCLR), presented at the ICML 2020 conference. Contrastive learning uses different clippings and variations of the same image to train a neural network until it learns robust representations to changes.
In their new work, Google’s research team used a variant of the SimCLR framework called Multi-Instance Contrastive Learning (MICLe), which learns stronger representations using multiple images of the same condition. This is often the case with medical datasets, where there are multiple images of the same patient, although the images may not be annotated for supervised learning.
“Unlabeled data is often available in large quantities in various medical settings. An important difference is that we use multiple views of the underlying pathology commonly found in medical imaging datasets to build image pairs for contrastive self-supervised learning, ”Azizi said.
When a self-supervised deep learning model is trained on different viewing angles of the same target, it learns multiple representations that are more robust to changes in point of view, imaging conditions, and other factors that could adversely affect its performance.
Put it all together
The self-supervised learning framework on Google researchers used three steps involved. First, the target neural network was trained on examples from the ImageNet dataset using SimCLR. Subsequently, the model was further trained using MICLe on a medical dataset that has multiple images for each patient. Finally, the model is fine-tuned to a limited dataset of images tagged for the target application.
Researchers tested the structure on two interpretation tasks of dermatology and chest X-ray. Compared to supervised pre-training, the self-supervised method provides a significant improvement in the accuracy, label efficiency and out-of-distribution generalization of medical imaging models, which is particularly important for clinical applications. Plus, it requires much less tagged data.
“Using self-supervised learning, we demonstrate that we can significantly reduce the need for expensive annotated data to build medical image classification models,” said Azizi. Specifically, in the dermatology business, they were able to train neural networks to match the performance of the baseline model using only one fifth of the annotated data.
“This will hopefully translate into significant time and cost savings in developing medical AI models. We hope this method will inspire explorations into new healthcare applications where acquiring annotated data has been challenging, ”said Azizi.
Ben Dickson is a software engineer and the founder of TechTalks. He writes about technology, business and politics.
This story originally appeared on Bdtechtalks.com. Copyright 2021
VentureBeat’s mission is to be a digital town square for technical decision makers to gain insights into transformative technology and transactions. Our site provides essential information on data technologies and strategies to guide you in driving your organizations. We invite you to become a member of our community, to access:
- updated information on topics of interest to you
- our newsletters
- thought-leading gated content and discounted access to our valuable events, such as Transform 2021: To know more
- network functions and more