Deep learning-based unlearning of dataset bias for MRI harmonisation and confound removal

Dinsdale, Nicola K.; Jenkinson, Mark; Namburete, Ana I. L.

doi:10.1016/j.neuroimage.2020.117689

Cited by 103 publications

(75 citation statements)

References 46 publications

(55 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Also, although PAC 2019 provides a true measurement for generalisability of models to unseen data (because the test set labels are hidden from the participants), this does not guarantee the generalisability to unseen scanning site (because the test set follows the same site and age distribution as the training set). For applications requiring site generalisability, see recent work aiming to address this specific issue ( 33 ).…”

Section: Discussionmentioning

confidence: 99%

Optimising a Simple Fully Convolutional Network for Accurate Brain Age Prediction in the PAC 2019 Challenge

et al. 2021

View full text Add to dashboard Cite

Brain age prediction from brain MRI scans not only helps improve brain ageing modelling generally, but also provides benchmarks for predictive analysis methods. Brain-age delta, which is the difference between a subject's predicted age and true age, has become a meaningful biomarker for the health of the brain. Here, we report the details of our brain age prediction models and results in the Predictive Analysis Challenge 2019. The aim of the challenge was to use T1-weighted brain MRIs to predict a subject's age in multicentre datasets. We apply a lightweight deep convolutional neural network architecture, Simple Fully Convolutional Neural Network (SFCN), and combined several techniques including data augmentation, transfer learning, model ensemble, and bias correction for brain age prediction. The model achieved first place in both of the two objectives in the PAC 2019 brain age prediction challenge: Mean absolute error (MAE) = 2.90 years without bias removal (Second Place = 3.09 yrs; Third Place = 3.33 yrs), and MAE = 2.95 years with bias removal, leading by a large margin (Second Place = 3.80 yrs; Third Place = 3.92 yrs).

show abstract

Section: Discussionmentioning

confidence: 99%

Optimising a Simple Fully Convolutional Network for Accurate Brain Age Prediction in the PAC 2019 Challenge

et al. 2021

View full text Add to dashboard Cite

show abstract

“…DANNs use a label predictor and a domain classifier to optimize the features to make the learned features discriminative for the main task but non-discriminative between the domains. Adapting the same framework as proposed in [186], Dinsdale et al [187] utilized an iterative update approach that aimed to Studies by Rozantsev et al [181] and Sun and Saenko [182] have adapted divergencebased approaches for domain adaptation by using a two-stream CNN architecture (one in the source domain with synthetic images and the other in the target domain with real images) with unshared weights and the DeepCORAL [183] architecture, respectively. Their methodologies provided a domain-invariant representation by trying to reduce the divergence (reduce the gap/distance) between feature distributions of source and target data distributions (both use non-medical images).…”

Section: Normalization Using Deep Learningmentioning

confidence: 99%

Making Radiomics More Reproducible across Scanner and Imaging Protocol Variations: A Review of Harmonization Methods

Mali

Ibrahim

Woodruff

et al. 2021

JPM

View full text Add to dashboard Cite

Radiomics converts medical images into mineable data via a high-throughput extraction of quantitative features used for clinical decision support. However, these radiomic features are susceptible to variation across scanners, acquisition protocols, and reconstruction settings. Various investigations have assessed the reproducibility and validation of radiomic features across these discrepancies. In this narrative review, we combine systematic keyword searches with prior domain knowledge to discuss various harmonization solutions to make the radiomic features more reproducible across various scanners and protocol settings. Different harmonization solutions are discussed and divided into two main categories: image domain and feature domain. The image domain category comprises methods such as the standardization of image acquisition, post-processing of raw sensor-level image data, data augmentation techniques, and style transfer. The feature domain category consists of methods such as the identification of reproducible features and normalization techniques such as statistical normalization, intensity harmonization, ComBat and its derivatives, and normalization using deep learning. We also reflect upon the importance of deep learning solutions for addressing variability across multi-centric radiomic studies especially using generative adversarial networks (GANs), neural style transfer (NST) techniques, or a combination of both. We cover a broader range of methods especially GANs and NST methods in more detail than previous reviews.

show abstract

“…The figures show example slices from the 3 canonical views per scan, as discussed above. (15,15,15) and (25,25,25).…”

Section: Image Pre-processingmentioning

confidence: 99%

“…Figure 2. 2D slice representation showing y-coronal, x-sagittal, and z-axial views of the brain (MNI template, OASIS scan) when for slices intersecting at(15,15,15) and(25,25,25).…”

mentioning

confidence: 99%

Alzheimer’s Disease Classification Accuracy is Improved by MRI Harmonization based on Attention-Guided Generative Adversarial Networks

Sinha

Thomopoulos

Lam

et al. 2021

Preprint

View full text Add to dashboard Cite

Alzheimer's disease (AD) accounts for 60% of dementia cases worldwide; patients with the disease typically suffer from irreversible memory loss and progressive decline in multiple cognitive domains. With brain imaging techniques such as magnetic resonance imaging (MRI), microscopic brain changes are detectable even before abnormal memory loss is detected clinically. Patterns of brain atrophy can be measured using MRI, which gives us an opportunity to facilitate AD detection using image classification techniques. Even so, MRI scanning protocols and scanners differ across studies. The resulting differences in image contrast and signal to noise make it important to train and test classification models on multiple datasets, and to handle shifts in image characteristics across protocols (also known as domain transfer or domain adaptation). Here, we examined whether adversarial domain adaptation can boost the performance of a Convolutional Neural Network (CNN) model designed to classify AD. To test this, we used an Attention-Guided Generative Adversarial Network (GAN) to harmonize images from three publicly available brain MRI datasets - ADNI, AIBL and OASIS - adjusting for scanner-dependent effects. Our AG-GAN optimized a joint objective function that included attention loss, pixel loss, cycle-consistency loss and adversarial loss; the model was trained bidirectionally in an end-to-end fashion. For AD classification, we adapted the popular 2D AlexNet CNN to handle 3D images. Classification based on harmonized MR images significantly outperformed classification based on the three datasets in non-harmonized form, motivating further work on image harmonization using adversarial techniques.

show abstract

Deep learning-based unlearning of dataset bias for MRI harmonisation and confound removal

Cited by 103 publications

References 46 publications

Optimising a Simple Fully Convolutional Network for Accurate Brain Age Prediction in the PAC 2019 Challenge

Optimising a Simple Fully Convolutional Network for Accurate Brain Age Prediction in the PAC 2019 Challenge

Making Radiomics More Reproducible across Scanner and Imaging Protocol Variations: A Review of Harmonization Methods

Alzheimer’s Disease Classification Accuracy is Improved by MRI Harmonization based on Attention-Guided Generative Adversarial Networks

Contact Info

Product

Resources

About