Connect with us

Tech

Self-supervised learning advances medical image classification

Published

on

 



Posted by Shekoofeh Azizi, AI Resident, Google Research

In recent years, there has been increasing interest in applying deep learning to medical imaging tasks, and exciting advances have been made in a variety of applications such as radiology, pathology, and dermatology. Despite interest, developing a medical image model remains difficult due to the often lack of high quality labeled data due to the time-consuming work required to annotate medical images. .. Given this, transfer learning is a common paradigm for building medical imaging models. In this approach, the model is first pretrained using supervised learning on a large labeled dataset (such as ImageNet), and then the learned general expressions are fine-tuned with medical data within the domain. increase.

Other more recent approaches that have proven successful in natural image recognition tasks, especially when there is a lack of labeled examples, are supervised contrasting pre-training followed by supervised learning. Use the fine-tuning of (SimCLR, MoCo, etc.). Pre-training with contrast learning learns general expressions by maximizing the match between different transformed views of the same image and minimizing the match between the transformed views of different images. Despite their success, these contrasting learning methods have received limited attention in medical image analysis and their effectiveness has not yet been investigated.

The “Big Self-Supervised Model Advanced Medical Image Classification” at the International Conference on Computer Vision (ICCV 2021) will study the effectiveness of self-supervised contrast learning as a pre-training strategy within the realm of medical image classification. .. We also propose multi-instance contrast learning (MICLe), a new approach to generalize contrast learning to take advantage of the special characteristics of medical image datasets. Experiment with two different medical image classification tasks: dermatological condition classification from digital camera images (27 categories) and multi-label chest x-ray classification (5 categories). We find that self-supervised learning with ImageNet, followed by additional self-supervised learning with unlabeled domain-specific medical images, significantly improves the accuracy of the medical image classifier. Specifically, it shows that self-supervised pre-training is superior to supervised pre-training, even when the full ImageNet dataset (14M images and 21.8K classes) is used for supervised pre-training. ..

SimCLR and Multi-instance Contrast Learning (MICLe) Our approach consists of three steps. (1) Self-monitoring pre-training of unlabeled natural images (using SimCLR). (2) Self-monitoring pre-training using unlabeled medical data (using either SimCLR or MICLe). Then, (3) task-specific supervised tweaks using labeled medical data.

Our approach consists of three steps: (1) self-monitoring pre-training with unlabeled ImageNet using SimCLR (2) additional self-monitoring pre-training with unlabeled medical images. When multiple images of each medical condition are available, a new multi-instance contrast learning (MICLe) strategy is used to build more informative positive pairs based on the different images. (3) Monitored fine-tuning of labeled medical images. Note that unlike step (1), steps (2) and (3) are specific to the task and dataset.

After the initial pre-training with SimCLR for unlabeled natural images is complete, the model is trained to capture the special characteristics of the medical image dataset. This can also be done with SimCLR, but this method builds positive pairs only by extension and does not easily leverage patient metadata for positive pair building. Alternatively, use MICLe. It builds a more informative positive pair for self-supervised learning, using multiple images of the underlying pathology of each patient’s case, when available. Such multi-instance data is often available in medical imaging datasets. For example, front and side views of the mammogram, retinal fundus images of each eye, and so on.

Given multiple images of a particular patient’s case, MICLe builds a positive pair of self-monitoring controlled learning by drawing two crops from two different images from the same patient’s case. Such images may be taken from different viewing angles and show different body parts with the same underlying medical condition. This provides a great opportunity for self-supervised learning algorithms to learn robust expressions in a direct way to viewpoint changes, imaging conditions, and other confounding factors. MICLe does not require class label information and relies only on different images of the underlying pathology, but its type may be unknown.

MICLe generalizes control learning and leverages the special characteristics of medical image datasets (patient metadata) to create realistic extensions and further improve image classifier performance.

Combining these self-supervised learning strategies, even in a highly competitive production environment, we achieved a significant improvement of 6.7% in top 1 accuracy in dermatological skin condition classification, with an average AUC of 1.1 in chest X-. Indicates that a% improvement can be achieved. Beyond the strong supervised baseline pre-trained in ray classification, ImageNet (a common protocol for training medical image analysis models). In addition, the self-monitoring model is robust to distribution shifts, demonstrating that it can be trained efficiently with a small number of labeled medical images.

Comparison of Supervised and Self-Trained Pre-Training Despite its simplicity, pre-training with MICLe is the original pre-training with SimCLR under the choice of different pre-training datasets and base network architectures. You will find that it consistently improves dermatological classification performance over training methods. Using MICLe for pre-training improves the accuracy of the top 1 dermatological classification by (1.18 ± 0.09)% over using SimCLR. The results show the benefits of leveraging additional metadata or domain knowledge to build more meaningful extensions for contrasting pre-training. In addition, our results show that the ResNet-152 (2x width) model outperforms the ResNet-50 (1x width) model or smaller models, so the wider and deeper the model, the better the performance. Indicates to do.

Comparison of supervised and self-supervised pre-training, followed by supervised fine-tuning using two architectures for dermatology and chest x-ray classification. Supervised learning utilizes unlabeled domain-specific medical images and far exceeds the pre-training of supervised ImageNet.

Improving generalization with a self-supervised model For each task, perform pre-training and fine-tuning using unlabeled and labeled data in the domain, respectively. It also uses different datasets from different clinical settings as shifted datasets to further evaluate the robustness of the method for out-of-domain data. For chest x-ray tasks, self-monitoring pre-training using either ImageNet or CheXpert data improves generalization, but keep in mind that stacking both will provide additional benefits. Also note that, as expected, using ImageNet only for self-monitoring pretraining will result in poor model performance compared to using only intradomain data for pretraining.

To test performance under distribution shifts, we provided additional labeled datasets for testing collected in different clinical settings for each task. You can see that the performance improvement of the distribution shift dataset (ChestX-ray14) by using self-monitoring pre-training (using both ImageNet and CheXpert data) is more significant than the original improvement of the CheXpert dataset. .. This is a valuable finding as generalization under a distribution shift is of paramount importance to clinical applications. In the dermatology task, similar trends are observed for another shifted dataset that was collected in a skin cancer clinic and had a high prevalence of malignant conditions. This shows that the robustness of self-supervised expressions for distribution shifts is consistent across tasks.

Evaluation of a model of a distribution shift dataset for a chest x-ray interpretation task. Use a model trained with intra-domain data to make predictions with additional shifted datasets (zero-shot transfer training) without further tweaking. Observe that self-monitored pre-training leads to a more robust and better expression for distribution shifts. Evaluation of a model of a distribution shift dataset for dermatology tasks. Our results suggest that self-monitoring pre-trained models are generally generalized to distribution shifts by MICLe pre-training for maximum benefit.

Improving Label Efficiency Further investigate the label efficiency of self-supervised models for medical image classification by fine-tuning the model in different parts of the labeled training data. Use label percentages in the range of 10% to 90% for both Derm and CheXpert training datasets and how performance changes using different label percentages available for dermatology tasks To find out. First, pre-training with the self-monitoring model can compensate for the inefficiency of labels in medical image classification, and the self-monitoring model is consistently above the surveillance baseline across the percentage of samples sampled. These results also suggest that reducing and fine-tuning the number of labeled cases results in a proportionally higher gain for MICLe. In fact, MICLe can use only 20% of the ResNet-50 (4x) training data and 30% of the ResNet152 (2x) training data to match the baseline.

Top 1 accuracy of dermatological condition classification of monitored models under MICLe, SimCLR, and various unlabeled pre-training datasets and label fractions of various sizes. MICLe can use only 20% of ResNet-50 (4x) training data to match baselines.

Conclusion Supervised pre-training of natural image datasets is commonly used to improve the classification of medical images. We investigated alternative strategies based on self-monitoring pre-training of unlabeled natural and medical images and found that supervised pre-training, the standard paradigm for training medical image analysis models, could be significantly improved. This approach is more accurate, label efficient, and can lead to a model that is robust to changing distributions. In addition, the proposed multi-instance contrast learning method (MICLe) can be used to create realistic extensions with additional metadata to further improve the performance of the image classifier.

Self-monitoring pre-training is much more scalable than supervised pre-training because it does not require class label annotations. We hope that this treatise will help popularize the use of self-monitoring approaches in medical image analysis and generate label-efficient and robust models suitable for large-scale clinical deployments in the real world.

Acknowledgments This work involved an interdisciplinary team of researchers, software engineers, clinicians, and transdisciplinary contributors across Google Health and Google Brain. Thanks to co-authors Basil Mustafa, Fiona Ryan, Zach Beaver, Jan Freyberg, Jon Deaton, Aaron Loh, Alan Karthikesalingam, Simon Kornblith, Ting Chen, Vivek Natarajan and Mohammad Norouzi. We would also like to thank Yuan Liu of Google Health for providing valuable feedback and our partners for accessing the dataset used in the study.

Sources

1/ https://Google.com/

2/ https://ai.googleblog.com/2021/10/self-supervised-learning-advances.html

The mention sources can contact us to remove/changing this article

What Are The Main Benefits Of Comparing Car Insurance Quotes Online

LOS ANGELES, CA / ACCESSWIRE / June 24, 2020, / Compare-autoinsurance.Org has launched a new blog post that presents the main benefits of comparing multiple car insurance quotes. For more info and free online quotes, please visit https://compare-autoinsurance.Org/the-advantages-of-comparing-prices-with-car-insurance-quotes-online/ The modern society has numerous technological advantages. One important advantage is the speed at which information is sent and received. With the help of the internet, the shopping habits of many persons have drastically changed. The car insurance industry hasn't remained untouched by these changes. On the internet, drivers can compare insurance prices and find out which sellers have the best offers. View photos The advantages of comparing online car insurance quotes are the following: Online quotes can be obtained from anywhere and at any time. Unlike physical insurance agencies, websites don't have a specific schedule and they are available at any time. Drivers that have busy working schedules, can compare quotes from anywhere and at any time, even at midnight. Multiple choices. Almost all insurance providers, no matter if they are well-known brands or just local insurers, have an online presence. Online quotes will allow policyholders the chance to discover multiple insurance companies and check their prices. Drivers are no longer required to get quotes from just a few known insurance companies. Also, local and regional insurers can provide lower insurance rates for the same services. Accurate insurance estimates. Online quotes can only be accurate if the customers provide accurate and real info about their car models and driving history. Lying about past driving incidents can make the price estimates to be lower, but when dealing with an insurance company lying to them is useless. Usually, insurance companies will do research about a potential customer before granting him coverage. Online quotes can be sorted easily. Although drivers are recommended to not choose a policy just based on its price, drivers can easily sort quotes by insurance price. Using brokerage websites will allow drivers to get quotes from multiple insurers, thus making the comparison faster and easier. For additional info, money-saving tips, and free car insurance quotes, visit https://compare-autoinsurance.Org/ Compare-autoinsurance.Org is an online provider of life, home, health, and auto insurance quotes. This website is unique because it does not simply stick to one kind of insurance provider, but brings the clients the best deals from many different online insurance carriers. In this way, clients have access to offers from multiple carriers all in one place: this website. On this site, customers have access to quotes for insurance plans from various agencies, such as local or nationwide agencies, brand names insurance companies, etc. "Online quotes can easily help drivers obtain better car insurance deals. All they have to do is to complete an online form with accurate and real info, then compare prices", said Russell Rabichev, Marketing Director of Internet Marketing Company. CONTACT: Company Name: Internet Marketing CompanyPerson for contact Name: Gurgu CPhone Number: (818) 359-3898Email: [email protected]: https://compare-autoinsurance.Org/ SOURCE: Compare-autoinsurance.Org View source version on accesswire.Com:https://www.Accesswire.Com/595055/What-Are-The-Main-Benefits-Of-Comparing-Car-Insurance-Quotes-Online View photos

ExBUlletin

to request, modification Contact us at Here or [email protected]