Simplified Transfer Learning for Chest Radiography Models Using Less                     Data

Sellergren, Andrew; Chen, Christina; Nabulsi, Zaid; Li, Yuanzhen; Maschinot, Aaron; Sarna, Aaron; Huang, Jenny; Lau, Charles T.; Kalidindi, Sreenivasa Raju; Etemadi, Mozziyar; Garcia-Vicente, Florencia; Melnick, David S.; Liu, Yun; Eswaran, Krishnan; Tse, Daniel; Beladia, Neeral; Krishnan, Dilip; Shetty, Shravya

doi:10.1148/radiol.212482

Cited by 24 publications

(7 citation statements)

References 14 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Although more research is necessary, self-supervised pretraining is a promising approach to image interpretation tasks, especially for tasks where datasets are small but contain clear, specific labels. A recent study found that pretraining the data on natural images, followed by pretraining on large weakly labeled CXR datasets, and finally, task-specific training on small labeled datasets could significantly reduce label requirements on the target dataset 42 ; thus the composition of large-scale out-of-domain pretraining and self-supervised pre training may be a fruitful direction for future work.…”

Section: Discussionmentioning

confidence: 99%

Self-Supervised Pretraining Enables High-Performance Chest X-Ray Interpretation Across Clinical Distributions

Gulati

Banerjee

et al. 2022

Preprint

View full text Add to dashboard Cite

Chest X-rays (CXRs) are a rich source of information for physicians, essential for disease diagnosis and treatment selection. Recent deep learning models aim to alleviate strain on medical resources and improve patient care by automating the detection of diseases from CXRs. However, shortages of labeled CXRs can pose a serious challenge when training models. Currently, models are generally pretrained on ImageNet, but they often need to then be finetuned on hundreds of thousands of labeled CXRs to achieve high performance. Therefore, the current approach to model development is not viable on tasks with only a small amount of labeled data. An emerging method for reducing reliance on large amounts of labeled data is self-supervised learning (SSL), which uses unlabeled CXR datasets to automatically learn features that can be leveraged for downstream interpretation tasks. In this work, we investigated whether self-supervised pretraining methods could outperform traditional ImageNet pretraining for chest X-ray interpretation. We found that SSL-pretrained models outperformed ImageNet-pretrained models on thirteen different datasets representing high diversity in geographies, clinical settings, and prediction tasks. We thus show that SSL on unlabeled CXR data is a promising pretraining approach for a wide variety of CXR interpretation tasks, enabling a shift away from costly labeled datasets.

show abstract

Section: Discussionmentioning

confidence: 99%

Self-Supervised Pretraining Enables High-Performance Chest X-Ray Interpretation Across Clinical Distributions

Gulati

Banerjee

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

Section: Methodsmentioning

confidence: 99%

“…The models used in this study were pretrained using contrastive self-supervised learning (CSL), which has been shown to perform better than ImageNet pretraining for CXR classification. 27 Specifically, an approach called MoCo-CXR 28 was used; it is an adaptation of the momentum contrast (MoCo) 29 approach for use with CXR data. MoCo maximizes the agreement between positive pairs of images (an image and augmentations thereof) while minimizing the agreement between negative pairs (any other pair of images).…”

Section: Model Developmentmentioning

confidence: 99%

Decision region analysis for generalizability of artificial intelligence models: estimating model generalizability in the case of cross-reactivity and population shift

Burgon,

Sahiner,

Petrick

et al. 2024

J. Med. Imag.

View full text Add to dashboard Cite

.PurposeUnderstanding an artificial intelligence (AI) model’s ability to generalize to its target population is critical to ensuring the safe and effective usage of AI in medical devices. A traditional generalizability assessment relies on the availability of large, diverse datasets, which are difficult to obtain in many medical imaging applications. We present an approach for enhanced generalizability assessment by examining the decision space beyond the available testing data distribution.ApproachVicinal distributions of virtual samples are generated by interpolating between triplets of test images. The generated virtual samples leverage the characteristics already in the test set, increasing the sample diversity while remaining close to the AI model’s data manifold. We demonstrate the generalizability assessment approach on the non-clinical tasks of classifying patient sex, race, COVID status, and age group from chest x-rays.ResultsDecision region composition analysis for generalizability indicated that a disproportionately large portion of the decision space belonged to a single “preferred” class for each task, despite comparable performance on the evaluation dataset. Evaluation using cross-reactivity and population shift strategies indicated a tendency to overpredict samples as belonging to the preferred class (e.g., COVID negative) for patients whose subgroup was not represented in the model development data.ConclusionsAn analysis of an AI model’s decision space has the potential to provide insight into model generalizability. Our approach uses the analysis of composition of the decision space to obtain an improved assessment of model generalizability in the case of limited test data.

show abstract

“…In this sense, Foundation Models training is an approach that trades off the need for task-specific data with the need for large amounts of data at pretraining. This is leveraged in hierarchical self-supervised pretraining which consists of a sequence of self-supervised training steps on decreasing amounts of increasingly task-relevant data, so as to tune the trade-off between data quantity and quality in ways that best match the data availability [10], [11].…”

Section: Data Requirementsmentioning

confidence: 99%

Towards Workflows for the Use of AI Foundation Models in Visual Inspection Applications

Rigotti,

Antognini,

Assaf

et al. 2023

ce papers

View full text Add to dashboard Cite

The latest successes in AI have been largely driven by a paradigm known as Foundation Models (FMs), large Neural Networks pretrained on massive datasets that thereby acquire impressive transfer learning capabilities to adapt to new tasks. The emerging properties of FMs have unlocked novel tantalizing applications for instance enabling the generation of fluent text and realistic images from text descriptions. The impact of FMs on technical domains like civil engineering is however still in its infancy, owing to a gap between research development and application use cases. This paper aims to help bridge this gap and promote adoption among technical practitioners, specifically in visual inspection applications for civil engineering. For that we analyze the requirements in terms of data availability making particular use cases amenable to the pretraining/fine‐tuning paradigm of FMs, i.e. situations where labeled data is scarce or costly, but unlabeled data is abundant. We then illustrate proof‐of‐concepts workflows using FMs, in visual inspection applications. We hope that our contribution will mark the start of conversations between AI researchers and civil engineers on the potential of FMs to accelerate workflows supporting vision tasks for maintenance inspections and decisions.

show abstract

Simplified Transfer Learning for Chest Radiography Models Using Less Data

Cited by 24 publications

References 14 publications

Self-Supervised Pretraining Enables High-Performance Chest X-Ray Interpretation Across Clinical Distributions

Self-Supervised Pretraining Enables High-Performance Chest X-Ray Interpretation Across Clinical Distributions

Decision region analysis for generalizability of artificial intelligence models: estimating model generalizability in the case of cross-reactivity and population shift

Towards Workflows for the Use of AI Foundation Models in Visual Inspection Applications

Contact Info

Product

Resources

About