CIOInsider India Magazine
Google's Self-Supervised AI Model Can Learn Multiple Representations of an Image in Medical Imaging
Versatility, high performance, high generalization capacity and multidisciplinary uses prove Deep Learning to be a promising solution in healthcare, especially around medical imaging. Most of the scientist community agree with this as medical imaging’s role is quite significant across different medical procedures such as early detection, monitoring, diagnosis, and treatment evaluation of various medical conditions.
In fact even the basic principles of deep learning can help the computer vision to understand medical image analysis. Along the line, deep learning has the ability to accelerate and enhance the accuracy of diagnosing health conditions in patients. Since in most cases medical experts often lack ample data that are clearly labelled in terms of what they are. Additionally there’s a shortage of labelled training and on the other hand, this training requires an expensive price to pay. These aspects induce hesitancy in deploying deep learning for many applications.
To counter these challenges, the AI community led by Shekoofeh Azizi at Google have suggested a technique called Multi-Instance Contrastive Learning (MICLe) that uses self-supervised learning to train deep learning models for medical imaging. This technique has already made news by showing outcomes such as reducing the need for annotated data, as well as improving the performance of deep learning models in medical imaging.
This Technique can Enable Neural Network that Learns Multiple Representations of a Single Image
Firstly, it is key to realize the fact that self-supervised machine learning is trained on the very basis of unlabeled data and can enable AI applications in areas that show difficulty in collecting clearly defined data, mainly in cases of treating deadly diseases like cancer.
Azizi and her team conducted experiments on two different tasks, one being dermatology skin condition classification from digital camera images and the other, multilabel chest X-ray classification. This was to showcase that self-supervised learning on ImageNet, followed by further self-supervised learning on unlabeled domain-specific medical images, significantly improved the accuracy of medical image classifiers.
How this turned out was that Azizi and her team employed MICLe to analyze several photos of a patient that did not have clearly labelled data points. To offer the algorithms an initial round of training, the first layer of the algorithm employed an available repository of photos with annotated data, in this instance ImageNet. Azizi's team then applied a second layer of photos, this time without tagged data, to force the system to form image pairs. The neural network was able to learn various representations of a single image as a result of this, which is crucial in medical research.
Following the aforementioned two layers of training, the researchers fine-tuned the algorithm for use on targets using a small data set of labelled images. According to the researchers, those algorithms showed that they can greatly cut the cost and time spent generating AI models for medical research, in addition to accuracy.
“On dermatology and chest X-ray classification, we improved top-1 accuracy by 6.7 percent and mean area under the curve (AUC) by 1.1 percent, respectively, beating strong supervised baselines pre-trained on ImageNet. Furthermore, we show that large self-supervised models are resistant to distribution shifts and may train quickly with a modest number of tagged medical images”, explained Azizi.
It is hoped that this would result in significant cost and time savings in the development of medical AI models. "We hope that this strategy will spur research into new health-care applications where obtaining annotated data has proven difficult," Azizi said
MICLe is built on Google's previous work on self-supervised convolutional neural network models. Google researchers presented Simple Framework for Contrastive Learning, or SimCLR, at the 2020 International Conference on Machine Learning (ICML), on which MICLe is based. SimCLR, simply put, learns many representations of the data it has by using multiple versions of the same image. As a result, the algorithm's identification became more robust and accurate.
A common factor in medical datasets is that there are several photos of the same patient present, even if the images are not labelled for supervised learning.
“In a range of medical domains, unlabeled data is frequently available in enormous amounts. One significant difference is that we create image pairings for contrastive self-supervised learning using different views of the underlying disease often found in medical imaging datasets”, Azizi explained.
Azizi says that, “we offer the unique MICLe approach, which constructs more informative positive pairs for self-supervised learning using multiple images of the underlying pathology per patient case”.
Since medical imaging cannot be controlled or coordinated, images in clinical therapies often contain various viewpoints and conditions. But AI has proven time and time again that it is capable of not only cutting down time, but beats issues in analyzing medical images.
Tackling Medical Imaging Issues through AI
Being a part of AI, Machine Learning (ML) can learn from data and make predictions or judgments based on that data without being explicitly programmed. ML employs three different learning methods: supervised, unsupervised, and semi-supervised learning. The extraction of features is one of the ML techniques, and choosing the right features for a particular problem necessitates the assistance of a domain expert.
Furthermore, the problem of feature selection is solved using deep learning (DL) techniques.
DL is a subset of machine learning, and it can automatically extract key characteristics from raw input data.
Basically, DL possesses two features which are, multiple processing layers that can learn various data features at several levels of abstraction, and unsupervised or supervised feature presentation learning on each layer.
DLA was taken from the field of computer vision and applied to medical image analysis in many forms. Supervised deep learning methods include recurrent neural networks (RNNs) and convolutional neural networks.
Coming to Azizi’s experimentation, Azizi said that, “we can considerably reduce the need for expensive annotated data to create medical image classification models using self-supervised learning”.
They were able to train the neural networks to match the baseline model performance on the dermatological experiment using only a fifth of the annotated data.
It is hoped that this would result in significant cost and time savings in the development of medical AI models. "We hope that this strategy will spur research into new health-care applications where obtaining annotated data has proven difficult," Azizi said.
