Corey Arnold

Minghui Zhang, Yangqian Wu, Hanxiao Zhang et al. · harvard

Open international challenges are becoming the de facto standard for assessing computer vision and image analysis algorithms. In recent years, new methods have extended the reach of pulmonary airway segmentation that is closer to the limit of image resolution. Since EXACT'09 pulmonary airway segmentation, limited effort has been directed to quantitative comparison of newly emerged algorithms driven by the maturity of deep learning based approaches and clinical drive for resolving finer details of distal airways for early intervention of pulmonary diseases. Thus far, public annotated datasets are extremely limited, hindering the development of data-driven methods and detailed performance evaluation of new algorithms. To provide a benchmark for the medical imaging community, we organized the Multi-site, Multi-domain Airway Tree Modeling (ATM'22), which was held as an official challenge event during the MICCAI 2022 conference. ATM'22 provides large-scale CT scans with detailed pulmonary airway annotation, including 500 CT scans (300 for training, 50 for validation, and 150 for testing). The dataset was collected from different sites and it further included a portion of noisy COVID-19 CTs with ground-glass opacity and consolidation. Twenty-three teams participated in the entire phase of the challenge and the algorithms for the top ten teams are reviewed in this paper. Quantitative and qualitative results revealed that deep learning models embedded with the topological continuity enhancement achieved superior performance in general. ATM'22 challenge holds as an open-call design, the training data and the gold standard evaluation are available upon successful registration via its homepage.

Vedrana Ivezić, Mara Pleasure, Ashwath Radhachandran et al.

Though self-supervised learning (SSL) has demonstrated incredible ability to learn robust representations from unlabeled data, the choice of optimal SSL strategy can lead to vastly different performance outcomes in specialized domains. Joint embedding architectures (JEAs) and joint embedding predictive architectures (JEPAs) have shown robustness to noise and strong semantic feature learning compared to pixel reconstruction-based SSL methods, leading to widespread adoption in medical imaging. However, no prior work has systematically investigated which SSL objective is better aligned with the spatial organization of clinically relevant signal. In this work, we empirically investigate how the choice of SSL method impacts the learned representations in medical imaging. We select two representative imaging modalities characterized by unique noise profiles: ultrasound and histopathology. When informative signal is spatially localized, as in histopathology, JEAs are more effective due to their view-invariance objective. In contrast, when diagnostically relevant information is globally structured, such as the macroscopic anatomy present in liver ultrasounds, JEPAs are optimal. These differences are especially evident in the clinical relevance of the learned features, as independently validated by board-certified radiologists and pathologists. Together, our results provide a framework for matching SSL objectives to the structural and noise properties of medical imaging modalities.

Saarang Panchavati, Uddhav Panchavati, Corey Arnold et al.

Electroencephalography (EEG) is a widely used tool for studying brain function, with applications in clinical neuroscience, diagnosis, and brain-computer interfaces (BCIs). Recent EEG foundation models trained on large unlabeled corpora aim to learn transferable representations, but their effectiveness remains unclear; reported improvements over smaller task-specific models are often modest, sensitive to downstream adaptation and fine-tuning strategies, and limited under linear probing. We hypothesize that one contributing factor is the reliance on signal reconstruction as the primary self-supervised learning (SSL) objective, which biases representations toward high-variance artifacts rather than task-relevant neural structure. To address this limitation, we explore an SSL paradigm based on Joint Embedding Predictive Architectures (JEPA), which learn by predicting latent representations instead of reconstructing raw signals. While earlier JEPA-style methods often rely on additional heuristics to ensure training stability, recent advances such as LeJEPA provide a more principled and stable formulation. We introduce Laya, the first EEG foundation model based on LeJEPA. Across a range of EEG benchmarks, Laya demonstrates improved performance under linear probing compared to reconstruction-based baselines, suggesting that latent predictive objectives offer a promising direction for learning transferable, high-level EEG representations.

Saarang Panchavati, Corey Arnold, William Speier

This work explores the potential of foundation models, specifically a Mamba-based selective state space model, for enhancing EEG analysis in neurological disorder diagnosis. EEG, crucial for diagnosing conditions like epilepsy, presents significant challenges due to its noisy, high-dimensional, and nonlinear nature. Traditional machine learning methods have made advances in automating EEG analysis but often fail to capture its complex spatio-temporal dynamics. Recent advances in deep learning, particularly in sequence modeling, offer new avenues for creating more generalized and expressive models capable of handling such complexities. By training a Mamba-based model on a large dataset containing seizure and non-seizure EEG recordings through a self-supervised reconstruction task followed by a seizure detection task, we demonstrate the model's effectiveness, achieving an AUROC of 0.72 on a held-out test set. This approach marks a significant step toward developing large-scale, clinically applicable foundation models for EEG data analysis.

Nithin Parthasarathy, James Soetedjo, Saarang Panchavati et al.

Amyotrophic lateral sclerosis (ALS) severely impairs patients' ability to communicate, often leading to a decline in their quality of life within a few years of diagnosis. The P300 speller brain-computer interface (BCI) offers an alternative communication method by interpreting a subject's EEG response to characters presented on a grid interface. This paper addresses the common speed limitations encountered in training efficient P300-based multi-subject classifiers by introducing innovative "across-subject" classifiers. We leverage a combination of the second-generation Generative Pre-Trained Transformer (GPT2) and Dijkstra's algorithm to optimize stimuli and suggest word completion choices based on typing history. Additionally, we employ a multi-layered smoothing technique to accommodate out-of-vocabulary (OOV) words. Through extensive simulations involving random sampling of EEG data from subjects, we demonstrate significant speed enhancements in typing passages containing rare and OOV words. These optimizations result in approximately 10% improvement in character-level typing speed and up to 40% improvement in multi-word prediction. We demonstrate that augmenting standard row/column highlighting techniques with layered word prediction yields close-to-optimal performance. Furthermore, we explore both "within-subject" and "across-subject" training techniques, showing that speed improvements are consistent across both approaches.

Ekaterina Redekop, Mara Pleasure, Vedrana Ivezic et al.

Foundation models in digital pathology use massive datasets to learn useful compact feature representations of complex histology images. However, there is limited transparency into what drives the correlation between dataset size and performance, raising the question of whether simply adding more data to increase performance is always necessary. In this study, we propose a prototype-guided diffusion model to generate high-fidelity synthetic pathology data at scale, enabling large-scale self-supervised learning and reducing reliance on real patient samples while preserving downstream performance. Using guidance from histological prototypes during sampling, our approach ensures biologically and diagnostically meaningful variations in the generated data. We demonstrate that self-supervised features trained on our synthetic dataset achieve competitive performance despite using ~60x-760x less data than models trained on large real-world datasets. Notably, models trained using our synthetic data showed statistically comparable or better performance across multiple evaluation metrics and tasks, even when compared to models trained on orders of magnitude larger datasets. Our hybrid approach, combining synthetic and real data, further enhanced performance, achieving top results in several evaluations. These findings underscore the potential of generative AI to create compelling training data for digital pathology, significantly reducing the reliance on extensive clinical datasets and highlighting the efficiency of our approach.

Irsyad Adam, Tengyue Zhang, Shrayes Raman et al.

Thyroid cancer is among the most common cancers in the United States. Thyroid nodules are frequently detected through ultrasound (US) imaging, and some require further evaluation via fine-needle aspiration (FNA) biopsy. Despite its effectiveness, FNA often leads to unnecessary biopsies of benign nodules, causing patient discomfort and anxiety. To address this, the American College of Radiology Thyroid Imaging Reporting and Data System (TI-RADS) has been developed to reduce benign biopsies. However, such systems are limited by interobserver variability. Recent deep learning approaches have sought to improve risk stratification, but they often fail to utilize the rich temporal and spatial context provided by US cine clips, which contain dynamic global information and surrounding structural changes across various views. In this work, we propose the Spatio-Temporal Cross Attention for Cine Thyroid Ultrasound Time Series Classification (STACT-Time) model, a novel representation learning framework that integrates imaging features from US cine clips with features from segmentation masks automatically generated by a pretrained model. By leveraging self-attention and cross-attention mechanisms, our model captures the rich temporal and spatial context of US cine clips while enhancing feature representation through segmentation-guided learning. Our model improves malignancy prediction compared to state-of-the-art models, achieving a cross-validation precision of 0.91 (plus or minus 0.02) and an F1 score of 0.89 (plus or minus 0.02). By reducing unnecessary biopsies of benign nodules while maintaining high sensitivity for malignancy detection, our model has the potential to enhance clinical decision-making and improve patient outcomes.

Shreeram Athreya, Carlos Olivares, Ameera Ismail et al.

Following successful large-vessel recanalization via endovascular thrombectomy (EVT) for acute ischemic stroke (AIS), some patients experience a complication known as no-reflow, defined by persistent microvascular hypoperfusion that undermines tissue recovery and worsens clinical outcomes. Although prompt identification is crucial, standard clinical practice relies on perfusion magnetic resonance imaging (MRI) within 24 hours post-procedure, delaying intervention. In this work, we introduce the first-ever machine learning (ML) framework to predict no-reflow immediately after EVT by leveraging previously unexplored intra-procedural digital subtraction angiography (DSA) sequences and clinical variables. Our retrospective analysis included AIS patients treated at UCLA Medical Center (2011-2024) who achieved favorable mTICI scores (2b-3) and underwent pre- and post-procedure MRI. No-reflow was defined as persistent hypoperfusion (Tmax > 6 s) on post-procedural imaging. From DSA sequences (AP and lateral views), we extracted statistical and temporal perfusion features from the target downstream territory to train ML classifiers for predicting no-reflow. Our novel method significantly outperformed a clinical-features baseline(AUC: 0.7703 $\pm$ 0.12 vs. 0.5728 $\pm$ 0.12; accuracy: 0.8125 $\pm$ 0.10 vs. 0.6331 $\pm$ 0.09), demonstrating that real-time DSA perfusion dynamics encode critical insights into microvascular integrity. This approach establishes a foundation for immediate, accurate no-reflow prediction, enabling clinicians to proactively manage high-risk patients without reliance on delayed imaging.

Vedrana Ivezić, Ashwath Radhachandran, Ekaterina Redekop et al.

Cytology is essential for cancer diagnostics and screening due to its minimally invasive nature. However, the development of robust deep learning models for digital cytology is challenging due to the heterogeneity in staining and preparation methods of samples, differences across organs, and the limited availability of large, diverse, annotated datasets. Developing a task-specific model for every cytology application is impractical and non-cytology-specific foundation models struggle to generalize to tasks in this domain where the emphasis is on cell morphology. To address these challenges, we introduce CytoFM, the first cytology self-supervised foundation model. Using iBOT, a self-supervised Vision Transformer (ViT) training framework incorporating masked image modeling and self-distillation, we pretrain CytoFM on a diverse collection of cytology datasets to learn robust, transferable representations. We evaluate CytoFM on multiple downstream cytology tasks, including breast cancer classification and cell type identification, using an attention-based multiple instance learning framework. Our results demonstrate that CytoFM performs better on two out of three downstream tasks than existing foundation models pretrained on histopathology (UNI) or natural images (iBOT-Imagenet). Visualizations of learned representations demonstrate our model is able to attend to cytologically relevant features. Despite a small pre-training dataset, CytoFM's promising results highlight the ability of task-agnostic pre-training approaches to learn robust and generalizable features from cytology data.

Nithin Parthasarathy, James Soetedjo, Saarang Panchavati et al.

Amyotrophic lateral sclerosis (ALS), a progressive neuromuscular degenerative disease, severely restricts patient communication capacity within a few years of onset, resulting in a significant deterioration of quality of life. The P300 speller brain computer interface (BCI) offers an alternative communication medium by leveraging a subject's EEG response to characters traditionally highlighted on a character grid on a graphical user interface (GUI). A recurring theme in P300-based research is enhancing performance to enable faster subject interaction. This study builds on that theme by addressing key limitations, particularly in the training of multi-subject classifiers, and by integrating advanced language models to optimize stimuli presentation and word prediction, thereby improving communication efficiency. Furthermore, various advanced large language models such as Generative Pre-Trained Transformer (GPT2), BERT, and BART, alongside Dijkstra's algorithm, are utilized to optimize stimuli and provide word completion choices based on the spelling history. In addition, a multi-layered smoothing approach is applied to allow for out-of-vocabulary (OOV) words. By conducting extensive simulations based on randomly sampled EEG data from subjects, we show substantial speed improvements in typing passages that include rare and out-of-vocabulary (OOV) words, with the extent of improvement varying depending on the language model utilized. The gains through such character-level interface optimizations are approximately 10%, and GPT2 for multi-word prediction provides gains of around 40%. In particular, some large language models achieve performance levels within 10% of the theoretical performance limits established in this study. In addition, both within and across subjects, training techniques are explored, and speed improvements are shown to hold in both cases.

Wenyuan Li, Zichen Wang, Jiayun Li et al.

Recently, semi-supervised learning methods based on generative adversarial networks (GANs) have received much attention. Among them, two distinct approaches have achieved competitive results on a variety of benchmark datasets. Bad GAN learns a classifier with unrealistic samples distributed on the complement of the support of the input data. Conversely, Triple GAN consists of a three-player game that tries to leverage good generated samples to boost classification results. In this paper, we perform a comprehensive comparison of these two approaches on different benchmark datasets. We demonstrate their different properties on image generation, and sensitivity to the amount of labeled data provided. By comprehensively comparing these two methods, we hope to shed light on the future of GAN-based semi-supervised learning.

Wenyuan Li, Yunlong Wang, Yong Cai et al.

Rare diseases affect a relatively small number of people, which limits investment in research for treatments and cures. Developing an efficient method for rare disease detection is a crucial first step towards subsequent clinical research. In this paper, we present a semi-supervised learning framework for rare disease detection using generative adversarial networks. Our method takes advantage of the large amount of unlabeled data for disease detection and achieves the best results in terms of precision-recall score compared to baseline techniques.

Justin Wood, Patrick Tan, Wei Wang et al.

A popular approach to topic modeling involves extracting co-occurring n-grams of a corpus into semantic themes. The set of n-grams in a theme represents an underlying topic, but most topic modeling approaches are not able to label these sets of words with a single n-gram. Such labels are useful for topic identification in summarization systems. This paper introduces a novel approach to labeling a group of n-grams comprising an individual topic. The approach taken is to complement the existing topic distributions over words with a known distribution based on a predefined set of topics. This is done by integrating existing labeled knowledge sources representing known potential topics into the probabilistic topic model. These knowledge sources are translated into a distribution and used to set the hyperparameters of the Dirichlet generated distribution over words. In the inference these modified distributions guide the convergence of the latent topics to conform with the complementary distributions. This approach ensures that the topic inference process is consistent with existing knowledge. The label assignment from the complementary knowledge sources are then transferred to the latent topics of the corpus. The results show both accurate label assignment to topics as well as improved topic generation than those obtained using various labeling approaches based off Latent Dirichlet allocation (LDA).