TransUNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers.

Medical image segmentation is crucial for healthcare, yet convolution-based methods like U-Net face limitations in modeling long-range dependencies. To address this, Transformers designed for sequence-to-sequence predictions have been integrated into medical image segmentation. However, a comprehensive understanding of Transformers' self-attention in U-Net components is lacking. TransUNet, first introduced in 2021, is widely recognized as one of the first models to integrate Transformer into medical image analysis. In this study, we present the versatile framework of TransUNet that encapsulates Transformers' self-attention into two key modules: (1) a Transformer encoder tokenizing image patches from a convolution neural network (CNN) feature map, facilitating global context extraction, and (2) a Transformer decoder refining candidate regions through cross-attention between proposals and U-Net features. These modules can be flexibly inserted into the U-Net backbone, resulting in three configurations: Encoder-only, Decoder-only, and Encoder+Decoder. TransUNet provides a library encompassing both 2D and 3D implementations, enabling users to easily tailor the chosen architecture. Our findings highlight the encoder's efficacy in modeling interactions among multiple abdominal organs and the decoder's strength in handling small targets like tumors. It excels in diverse medical applications, such as multi-organ segmentation, pancreatic tumor segmentation, and hepatic vessel segmentation. Notably, our TransUNet achieves a significant average Dice improvement of 1.06% and 4.30% for multi-organ segmentation and pancreatic tumor segmentation, respectively, when compared to the highly competitive nn-UNet, and surpasses the top-1 solution in the BrasTS2021 challenge. 2D/3D Code and models are available at https://github.com/Beckschen/TransUNet and https://github.com/Beckschen/TransUNet-3D, respectively.

Vision-based estimation of fatigue and engagement in cognitive training sessions.

Computerized cognitive training (CCT) is a scalable, well-tolerated intervention that has promise for slowing cognitive decline. The effectiveness of CCT is often affected by a lack of effective engagement. Mental fatigue is a the primary factor for compromising effective engagement in CCT, particularly in older adults at risk for dementia. There is a need for scalable, automated measures that can constantly monitor and reliably detect mental fatigue during CCT. Here, we develop and validate a novel Recurrent Video Transformer (RVT) method for monitoring real-time mental fatigue in older adults with mild cognitive impairment using their video-recorded facial gestures during CCT. The RVT model achieved the highest balanced accuracy (79.58%) and precision (0.82) compared to the prior models for binary and multi-class classification of mental fatigue. We also validated our model by significantly relating to reaction time across CCT tasks (Waldχ2=5.16,p=0.023). By leveraging dynamic temporal information, the RVT model demonstrates the potential to accurately measure real-time mental fatigue, laying the foundation for future CCT research aiming to enhance effective engagement by timely prevention of mental fatigue.

Federated learning with knowledge distillation for multi-organ segmentation with partially labeled datasets.

The state-of-the-art multi-organ CT segmentation relies on deep learning models, which only generalize when trained on large samples of carefully curated data. However, it is challenging to train a single model that can segment all organs and types of tumors since most large datasets are partially labeled or are acquired across multiple institutes that may differ in their acquisitions. A possible solution is Federated learning, which is often used to train models on multi-institutional datasets where the data is not shared across sites. However, predictions of federated learning can be unreliable after the model is locally updated at sites due to 'catastrophic forgetting'. Here, we address this issue by using knowledge distillation (KD) so that the local training is regularized with the knowledge of a global model and pre-trained organ-specific segmentation models. We implement the models in a multi-head U-Net architecture that learns a shared embedding space for different organ segmentation, thereby obtaining multi-organ predictions without repeated processes. We evaluate the proposed method using 8 publicly available abdominal CT datasets of 7 different organs. Of those datasets, 889 CTs were used for training, 233 for internal testing, and 30 volumes for external testing. Experimental results verified that our proposed method substantially outperforms other state-of-the-art methods in terms of accuracy, inference time, and the number of parameters.

LSOR: Longitudinally-Consistent Self-Organized Representation Learning.

Interpretability is a key issue when applying deep learning models to longitudinal brain MRIs. One way to address this issue is by visualizing the high-dimensional latent spaces generated by deep learning via self-organizing maps (SOM). SOM separates the latent space into clusters and then maps the cluster centers to a discrete (typically 2D) grid preserving the high-dimensional relationship between clusters. However, learning SOM in a high-dimensional latent space tends to be unstable, especially in a self-supervision setting. Furthermore, the learned SOM grid does not necessarily capture clinically interesting information, such as brain age. To resolve these issues, we propose the first self-supervised SOM approach that derives a high-dimensional, interpretable representation stratified by brain age solely based on longitudinal brain MRIs (i.e., without demographic or cognitive information). Called Longitudinally-consistent Self-Organized Representation learning (LSOR), the method is stable during training as it relies on soft clustering (vs. the hard cluster assignments used by existing SOM). Furthermore, our approach generates a latent space stratified according to brain age by aligning trajectories inferred from longitudinal MRIs to the reference vector associated with the corresponding SOM cluster. When applied to longitudinal MRIs of the Alzheimer's Disease Neuroimaging Initiative (ADNI, N=632), LSOR generates an interpretable latent space and achieves comparable or higher accuracy than the state-of-the-art representations with respect to the downstream tasks of classification (static vs. progressive mild cognitive impairment) and regression (determining ADAS-Cog score of all subjects). The code is available at https://github.com/ouyangjiahong/longitudinal-som-single-modality.

An Explainable Geometric-Weighted Graph Attention Network for Identifying Functional Networks Associated with Gait Impairment.

One of the hallmark symptoms of Parkinson's Disease (PD) is the progressive loss of postural reflexes, which eventually leads to gait difficulties and balance problems. Identifying disruptions in brain function associated with gait impairment could be crucial in better understanding PD motor progression, thus advancing the development of more effective and personalized therapeutics. In this work, we present an explainable, geometric, weighted-graph attention neural network (xGW-GAT) to identify functional networks predictive of the progression of gait difficulties in individuals with PD. xGW-GAT predicts the multi-class gait impairment on the MDS-Unified PD Rating Scale (MDS-UPDRS). Our computational- and data-efficient model represents functional connectomes as symmetric positive definite (SPD) matrices on a Riemannian manifold to explicitly encode pairwise interactions of entire connectomes, based on which we learn an attention mask yielding individual- and group-level explainability. Applied to our resting-state functional MRI (rs-fMRI) dataset of individuals with PD, xGW-GAT identifies functional connectivity patterns associated with gait impairment in PD and offers interpretable explanations of functional subnetworks associated with motor impairment. Our model successfully outperforms several existing methods while simultaneously revealing clinically-relevant connectivity patterns. The source code is available at https://github.com/favour-nerrise/xGW-GAT.

An Explainable Geometric-Weighted Graph Attention Network for Identifying Functional Networks Associated with Gait Impairment.

One of the hallmark symptoms of Parkinson's Disease (PD) is the progressive loss of postural reflexes, which eventually leads to gait difficulties and balance problems. Identifying disruptions in brain function associated with gait impairment could be crucial in better understanding PD motor progression, thus advancing the development of more effective and personalized therapeutics. In this work, we present an explainable, geometric, weighted-graph attention neural network (xGW-GAT) to identify functional networks predictive of the progression of gait difficulties in individuals with PD. xGW-GAT predicts the multi-class gait impairment on the MDS-Unified PD Rating Scale (MDS-UPDRS). Our computational- and data-efficient model represents functional connectomes as symmetric positive definite (SPD) matrices on a Riemannian manifold to explicitly encode pairwise interactions of entire connectomes, based on which we learn an attention mask yielding individual- and group-level explain-ability. Applied to our resting-state functional MRI (rs-fMRI) dataset of individuals with PD, xGW-GAT identifies functional connectivity patterns associated with gait impairment in PD and offers interpretable explanations of functional subnetworks associated with motor impairment. Our model successfully outperforms several existing methods while simultaneously revealing clinically-relevant connectivity patterns. The source code is available at https://github.com/favour-nerrise/xGW-GAT.

CCA identifies a neurophysiological marker of adaptation capacity that is reliably linked to internal locus of control of cognition in amnestic MCI.

Locus of control (LOC) describes whether an individual thinks that they themselves (internal LOC) or external factors (external LOC) have more influence on their lives. LOC varies by domain, and a person's LOC for their intellectual capacities (LOC-Cognition) may be a marker of resilience in older adults at risk for dementia, with internal LOC-Cognition relating to better outcomes and improved treatment adherence. Vagal control, a key component of parasympathetic autonomic nervous system (ANS) regulation, may reflect a neurophysiological biomarker of internal LOC-Cognition. We used canonical correlation analysis (CCA) to identify a shared neurophysiological marker of ANS regulation from electrocardiogram (during auditory working memory) and functional connectivity (FC) data. A canonical variable from root mean square of successive differences (RMSSD) time series and between-network FC was significantly related to internal LOC-Cognition (ß = 0.266, SE = 0.971, CI = [0.190, 4.073], p = 0.031) in 65 participants (mean age = 74.7, 32 female) with amnestic mild cognitive impairment (aMCI). Follow-up data from 55 of these individuals (mean age = 73.6, 22 females) was used to show reliability of this relationship (ß = 0.271, SE = 0.971, CI = [0.033, 2.630], p = 0.047), and a second sample (40 participants with aMCI/healthy cognition, mean age = 72.7, 24 females) showed that the canonical vector biomarker generalized to visual working memory (ß = 0.36, SE = 0.136, CI = [0.023, 0.574], p = 0.037), but not inhibition task RMSSD data (ß = 0.08, SE = 1.486, CI = [- 0.354, 0.657], p = 0.685). This canonical vector may represent a biomarker of autonomic regulation that explains how some older adults maintain internal LOC-Cognition as dementia progresses. Future work should further test the causality of this relationship and the modifiability of this biomarker.

Generating Realistic Brain MRIs via a Conditional Diffusion Probabilistic Model.

As acquiring MRIs is expensive, neuroscience studies struggle to attain a sufficient number of them for properly training deep learning models. This challenge could be reduced by MRI synthesis, for which Generative Adversarial Networks (GANs) are popular. GANs, however, are commonly unstable and struggle with creating diverse and high-quality data. A more stable alternative is Diffusion Probabilistic Models (DPMs) with a fine-grained training strategy. To overcome their need for extensive computational resources, we propose a conditional DPM (cDPM) with a memory-efficient process that generates realistic-looking brain MRIs. To this end, we train a 2D cDPM to generate an MRI subvolume conditioned on another subset of slices from the same MRI. By generating slices using arbitrary combinations between condition and target slices, the model only requires limited computational resources to learn interdependencies between slices even if they are spatially far apart. After having learned these dependencies via an attention network, a new anatomy-consistent 3D brain MRI is generated by repeatedly applying the cDPM. Our experiments demonstrate that our method can generate high-quality 3D MRIs that share a similar distribution to real MRIs while still diversifying the training set. The code is available at https://github.com/xiaoiker/mask3DMRI_diffusion and also will be released as part of MONAI, at https://github.com/Project-MONAI/GenerativeModels.

One-shot Federated Learning on Medical Data using Knowledge Distillation with Image Synthesis and Client Model Adaptation.

One-shot federated learning (FL) has emerged as a promising solution in scenarios where multiple communication rounds are not practical. Notably, as feature distributions in medical data are less discriminative than those of natural images, robust global model training with FL is non-trivial and can lead to overfitting. To address this issue, we propose a novel one-shot FL framework leveraging Image Synthesis and Client model Adaptation (FedISCA) with knowledge distillation (KD). To prevent overfitting, we generate diverse synthetic images ranging from random noise to realistic images. This approach (i) alleviates data privacy concerns and (ii) facilitates robust global model training using KD with decentralized client models. To mitigate domain disparity in the early stages of synthesis, we design noise-adapted client models where batch normalization statistics on random noise (synthetic images) are updated to enhance KD. Lastly, the global model is trained with both the original and noise-adapted client models via KD and synthetic images. This process is repeated till global model convergence. Extensive evaluation of this design on five small- and three large-scale medical image classification datasets reveals superior accuracy over prior methods. Code is available at https://github.com/myeongkyunkang/FedISCA.

A Penalty Approach for Normalizing Feature Distributions to Build Confounder-Free Models.

Translating the use of modern machine learning algorithms into clinical applications requires settling challenges related to explain-ability and management of nuanced confounding factors. To suitably interpret the results, removing or explaining the effect of confounding variables (or metadata) is essential. Confounding variables affect the relationship between input training data and target outputs. Accordingly, when we train a model on such data, confounding variables will bias the distribution of the learned features. A recent promising solution, Meta-Data Normalization (MDN), estimates the linear relationship between the metadata and each feature based on a non-trainable closed-form solution. However, this estimation is confined by the sample size of a mini-batch and thereby may result in an oscillating performance. In this paper, we extend the MDN method by applying a Penalty approach (referred to as PDMN). We cast the problem into a bi-level nested optimization problem. We then approximate that objective using a penalty method so that the linear parameters within the MDN layer are trainable and learned on all samples. This enables PMDN to be plugged into any architectures, even those unfit to run batch-level operations such as transformers and recurrent models. We show improvement in model accuracy and independence from the confounders using PMDN over MDN in a synthetic experiment and a multi-label, multi-site classification of magnetic resonance images.

Multiple Instance Neuroimage Transformer.

For the first time, we propose using a multiple instance learning based convolution-free transformer model, called Multiple Instance Neuroimage Transformer (MINiT), for the classification of T1-weighted (T1w) MRIs. We first present several variants of transformer models adopted for neuroimages. These models extract non-overlapping 3D blocks from the input volume and perform multi-headed self-attention on a sequence of their linear projections. MINiT, on the other hand, treats each of the non-overlapping 3D blocks of the input MRI as its own instance, splitting it further into non-overlapping 3D patches, on which multi-headed self-attention is computed. As a proof-of-concept, we evaluate the efficacy of our model by training it to identify sex from T1w-MRIs of two public datasets: Adolescent Brain Cognitive Development (ABCD) and the National Consortium on Alcohol and Neurodevelopment in Adolescence (NCANDA). The learned attention maps highlight voxels contributing to identifying sex differences in brain morphometry. The code is available at https://github.com/singlaayush/MINIT.

Joint Graph Convolution for Analyzing Brain Structural and Functional Connectome.

The white-matter (micro-)structural architecture of the brain promotes synchrony among neuronal populations, giving rise to richly patterned functional connections. A fundamental problem for systems neuroscience is determining the best way to relate structural and functional networks quantified by diffusion tensor imaging and resting-state functional MRI. As one of the state-of-the-art approaches for network analysis, graph convolutional networks (GCN) have been separately used to analyze functional and structural networks, but have not been applied to explore inter-network relationships. In this work, we propose to couple the two networks of an individual by adding inter-network edges between corresponding brain regions, so that the joint structure-function graph can be directly analyzed by a single GCN. The weights of inter-network edges are learnable, reflecting non-uniform structure-function coupling strength across the brain. We apply our Joint-GCN to predict age and sex of 662 participants from the public dataset of the National Consortium on Alcohol and Neurodevelopment in Adolescence (NCANDA) based on their functional and micro-structural white-matter networks. Our results support that the proposed Joint-GCN outperforms existing multi-modal graph learning approaches for analyzing structural and functional networks.

GaitForeMer: Self-Supervised Pre-Training of Transformers via Human Motion Forecasting for Few-Shot Gait Impairment Severity Estimation.

Parkinson's disease (PD) is a neurological disorder that has a variety of observable motor-related symptoms such as slow movement, tremor, muscular rigidity, and impaired posture. PD is typically diagnosed by evaluating the severity of motor impairments according to scoring systems such as the Movement Disorder Society Unified Parkinson's Disease Rating Scale (MDS-UPDRS). Automated severity prediction using video recordings of individuals provides a promising route for non-intrusive monitoring of motor impairments. However, the limited size of PD gait data hinders model ability and clinical potential. Because of this clinical data scarcity and inspired by the recent advances in self-supervised large-scale language models like GPT-3, we use human motion forecasting as an effective self-supervised pre-training task for the estimation of motor impairment severity. We introduce GaitForeMer, Gait Forecasting and impairment estimation transforMer, which is first pre-trained on public datasets to forecast gait movements and then applied to clinical data to predict MDS-UPDRS gait impairment severity. Our method outperforms previous approaches that rely solely on clinical data by a large margin, achieving an F1 score of 0.76, precision of 0.79, and recall of 0.75. Using GaitForeMer, we show how public human movement data repositories can assist clinical use cases through learning universal motion representations. The code is available at https://github.com/markendo/GaitForeMer.

Bridging the Gap between Deep Learning and Hypothesis-Driven Analysis via Permutation Testing.

A fundamental approach in neuroscience research is to test hypotheses based on neuropsychological and behavioral measures, i.e., whether certain factors (e.g., related to life events) are associated with an outcome (e.g., depression). In recent years, deep learning has become a potential alternative approach for conducting such analyses by predicting an outcome from a collection of factors and identifying the most "informative" ones driving the prediction. However, this approach has had limited impact as its findings are not linked to statistical significance of factors supporting hypotheses. In this article, we proposed a flexible and scalable approach based on the concept of permutation testing that integrates hypothesis testing into the data-driven deep learning analysis. We apply our approach to the yearly self-reported assessments of 621 adolescent participants of the National Consortium of Alcohol and Neurodevelopment in Adolescence (NCANDA) to predict negative valence, a symptom of major depressive disorder according to the NIMH Research Domain Criteria (RDoC). Our method successfully identifies categories of risk factors that further explain the symptom.

Semantic instance segmentation with discriminative deep supervision for medical images.

Semantic instance segmentation is crucial for many medical image analysis applications, including computational pathology and automated radiation therapy. Existing methods for this task can be roughly classified into two categories: (1) proposal-based methods and (2) proposal-free methods. However, in medical images, the irregular shape-variations and crowding instances (e.g., nuclei and cells) make it hard for the proposal-based methods to achieve robust instance localization. On the other hand, ambiguous boundaries caused by the low-contrast nature of medical images (e.g., CT images) challenge the accuracy of the proposal-free methods. To tackle these issues, we propose a proposal-free segmentation network with discriminative deep supervision (DDS), which at the same time allows us to gain the power of the proposal-based method. The DDS module is interleaved with a carefully designed proposal-free segmentation backbone in our network. Consequently, the features learned by the backbone network become more sensitive to instance localization. Also, with the proposed DDS module, robust pixel-wise instance-level cues (especially structural information) are introduced for semantic segmentation. Extensive experiments on three datasets, i.e., a nuclei dataset, a pelvic CT image dataset, and a synthetic dataset, demonstrate the superior performance of the proposed algorithm compared to the previous works.

Self-supervised learning of neighborhood embedding for longitudinal MRI.

In recent years, several deep learning models recommend first to represent Magnetic Resonance Imaging (MRI) as latent features before performing a downstream task of interest (such as classification or regression). The performance of the downstream task generally improves when these latent representations are explicitly associated with factors of interest. For example, we derived such a representation for capturing brain aging by applying self-supervised learning to longitudinal MRIs and then used the resulting encoding to automatically identify diseases accelerating the aging of the brain. We now propose a refinement of this representation by replacing the linear modeling of brain aging with one that is consistent in local neighborhoods in the latent space. Called Longitudinal Neighborhood Embedding (LNE), we derive an encoding so that neighborhoods are age-consistent (i.e., brain MRIs of different subjects with similar brain ages are in close proximity of each other) and progression-consistent, i.e., the latent space is defined by a smooth trajectory field where each trajectory captures changes in brain ages between a pair of MRIs extracted from a longitudinal sequence. To make the problem computationally tractable, we further propose a strategy for mini-batch sampling so that the resulting local neighborhoods accurately approximate the ones that would be defined based on the whole cohort. We evaluate LNE on three different downstream tasks: (1) to predict chronological age from T1-w MRI of 274 healthy subjects participating in a study at SRI International; (2) to distinguish Normal Control (NC) from Alzheimer's Disease (AD) and stable Mild Cognitive Impairment (sMCI) from progressive Mild Cognitive Impairment (pMCI) based on T1-w MRI of 632 participants of the Alzheimer's Disease Neuroimaging Initiative (ADNI); and (3) to distinguish no-to-low from moderate-to-heavy alcohol drinkers based on fractional anisotropy derived from diffusion tensor MRIs of 764 adolescents recruited by the National Consortium on Alcohol and NeuroDevelopment in Adolescence (NCANDA). Across the three data sets, the visualization of the smooth trajectory vector fields and superior accuracy on downstream tasks demonstrate the strength of the proposed method over existing self-supervised methods in extracting information related to brain aging, which could help study the impact of substance use and neurodegenerative disorders. The code is available at https://github.com/ouyangjiahong/longitudinal-neighbourhood-embedding.

Detecting negative valence symptoms in adolescents based on longitudinal self-reports and behavioral assessments.

BACKGROUND: Given the high prevalence of depressive symptoms reported by adolescents and associated risk of experiencing psychiatric disorders as adults, differentiating the trajectories of the symptoms related to negative valence at an individual level could be crucial in gaining a better understanding of their effects later in life. METHODS: A longitudinal deep learning framework is presented, identifying self-reported and behavioral measurements that detect the depressive symptoms associated with the Negative Valence System domain of the NIMH Research Domain Criteria (RDoC). RESULTS: Applied to the annual records of 621 participants (age range: 12 to 17 years) of the National Consortium on Alcohol and Neurodevelopment in Adolescence (NCANDA), the deep learning framework identifies predictors of negative valence symptoms, which include lower extraversion, poorer sleep quality, impaired executive control function and factors related to substance use. LIMITATIONS: The results rely mainly on self-reported measures and do not provide information about the underlying neural correlates. Also, a larger sample is required to understand the role of sex and other demographics related to the risk of experiencing symptoms of negative valence. CONCLUSIONS: These results provide new information about predictors of negative valence symptoms in individuals during adolescence that could be critical in understanding the development of depression and identifying targets for intervention. Importantly, findings can inform preventive and treatment approaches for depression in adolescents, focusing on a unique predictor set of modifiable modulators to include factors such as sleep hygiene training, cognitive-emotional therapy enhancing coping and controllability experience and/or substance use interventions.

A Novel Explainability Approach for Technology-Driven Translational Research on Brain Aging.

Brain aging leads to difficulties in functional independence. Mitigating these difficulties can benefit from technology that predicts, monitors, and modifies brain aging. Translational research prioritizes solutions that can be causally linked to specific pathophysiologies at the same time as demonstrating improvements in impactful real-world outcome measures. This poses a challenge for brain aging technology that needs to address the tension between mechanism-driven precision and clinical relevance. In the current opinion, by synthesizing emerging mechanistic, translational, and clinical research-related frameworks, and our own development of technology-driven brain aging research, we suggest incorporating the appreciation of four desiderata (causality, informativeness, transferability, and fairness) of explainability into early-stage research that designs and tests brain aging technology. We apply a series of work on electrocardiography-based "peripheral" neuroplasticity markers from our work as an illustration of our proposed approach. We believe this novel approach will promote the development and adoption of brain aging technology that links and addresses brain pathophysiology and functional independence in the field of translational research.

Disentangling Normal Aging From Severity of Disease via Weak Supervision on Longitudinal MRI.

The continuous progression of neurological diseases are often categorized into conditions according to their severity. To relate the severity to changes in brain morphometry, there is a growing interest in replacing these categories with a continuous severity scale that longitudinal MRIs are mapped onto via deep learning algorithms. However, existing methods based on supervised learning require large numbers of samples and those that do not, such as self-supervised models, fail to clearly separate the disease effect from normal aging. Here, we propose to explicitly disentangle those two factors via weak-supervision. In other words, training is based on longitudinal MRIs being labelled either normal or diseased so that the training data can be augmented with samples from disease categories that are not of primary interest to the analysis. We do so by encouraging trajectories of controls to be fully encoded by the direction associated with brain aging. Furthermore, an orthogonal direction linked to disease severity captures the residual component from normal aging in the diseased cohort. Hence, the proposed method quantifies disease severity and its progression speed in individuals without knowing their condition. We apply the proposed method on data from the Alzheimer's Disease Neuroimaging Initiative (ADNI, N =632 ). We then show that the model properly disentangled normal aging from the severity of cognitive impairment by plotting the resulting disentangled factors of each subject and generating simulated MRIs for a given chronological age and condition. Moreover, our representation obtains higher balanced accuracy when used for two downstream classification tasks compared to other pre-training approaches. The code for our weak-supervised approach is available at https://github.com/ouyangjiahong/longitudinal-direction-disentangle.

Rethinking Architecture Design for Tackling Data Heterogeneity in Federated Learning.

Federated learning is an emerging research paradigm enabling collaborative training of machine learning models among different organizations while keeping data private at each institution. Despite recent progress, there remain fundamental challenges such as the lack of convergence and the potential for catastrophic forgetting across real-world heterogeneous devices. In this paper, we demonstrate that self-attention-based architectures (e.g., Transformers) are more robust to distribution shifts and hence improve federated learning over heterogeneous data. Concretely, we conduct the first rigorous empirical investigation of different neural architectures across a range of federated algorithms, real-world benchmarks, and heterogeneous data splits. Our experiments show that simply replacing convolutional networks with Transformers can greatly reduce catastrophic forgetting of previous devices, accelerate convergence, and reach a better global model, especially when dealing with heterogeneous data. We release our code and pretrained models to encourage future exploration in robust architectures as an alternative to current research efforts on the optimization front.