Vedrana Ivezić

Ashwath Radhachandran, Vedrana Ivezić, Shreeram Athreya et al.

Ultrasound (US) imaging poses unique challenges for representation learning due to its inherently noisy acquisition process. The low signal-to-noise ratio and stochastic speckle patterns hinder standard self-supervised learning methods relying on a pixel-level reconstruction objective. Joint-Embedding Predictive Architectures (JEPAs) address this drawback by predicting masked latent representations rather than raw pixels. However, standard approaches depend on hyperparameter-brittle and computationally expensive online teachers updated via exponential moving average. We propose US-JEPA, a self-supervised framework that adopts the Static-teacher Asymmetric Latent Training (SALT) objective. By using a frozen, domain-specific teacher to provide stable latent targets, US-JEPA decouples student-teacher optimization and pushes the student to expand upon the semantic priors of the teacher. In addition, we provide the first rigorous comparison of all publicly available state-of-the-art ultrasound foundation models on UltraBench, a public dataset benchmark spanning multiple organs and pathological conditions. Under linear probing for diverse classification tasks, US-JEPA achieves performance competitive with or superior to domain-specific and universal vision foundation model baselines. Our results demonstrate that masked latent prediction provides a stable and efficient path toward robust ultrasound representations.

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.

Shreeram Athreya, Andrew Melehy, Sujit Silas Armstrong Suthahar et al.

Objective: Molecular testing (MT) classifies cytologically indeterminate thyroid nodules as benign or malignant with high sensitivity but low positive predictive value (PPV), only using molecular profiles, ignoring ultrasound (US) imaging and biopsy. We address this limitation by applying attention multiple instance learning (AMIL) to US images. Methods: We retrospectively reviewed 333 patients with indeterminate thyroid nodules at UCLA medical center (259 benign, 74 malignant). A multi-modal deep learning AMIL model was developed, combining US images and MT to classify the nodules as benign or malignant and enhance the malignancy risk stratification of MT. Results: The final AMIL model matched MT sensitivity (0.946) while significantly improving PPV (0.477 vs 0.448 for MT alone), indicating fewer false positives while maintaining high sensitivity. Conclusion: Our approach reduces false positives compared to MT while maintaining the same ability to identify positive cases, potentially reducing unnecessary benign thyroid resections in patients with indeterminate nodules.

Shreeram Athreya, Ashwath Radhachandran, Vedrana Ivezić et al.

Purpose: The objective of this work is to introduce an advanced framework designed to enhance ultrasound images, especially those captured by portable hand-held devices, which often produce lower quality images due to hardware constraints. Additionally, this framework is uniquely capable of effectively handling non-registered input ultrasound image pairs, addressing a common challenge in medical imaging. Materials and Methods: In this retrospective study, we utilized an enhanced generative adversarial network (CycleGAN) model for ultrasound image enhancement across five organ systems. Perceptual loss, derived from deep features of pretrained neural networks, is applied to ensure the human-perceptual quality of the enhanced images. These images are compared with paired images acquired from high resolution devices to demonstrate the model's ability to generate realistic high-quality images across organ systems. Results: Preliminary validation of the framework reveals promising performance metrics. The model generates images that result in a Structural Similarity Index (SSI) score of 0.722, Locally Normalized Cross-Correlation (LNCC) score of 0.902 and 28.802 for the Peak Signal-to-Noise Ratio (PSNR) metric. Conclusion: This work presents a significant advancement in medical imaging through the development of a CycleGAN model enhanced with Perceptual Loss (PL), effectively bridging the quality gap between ultrasound images from varied devices. By training on paired images, the model not only improves image quality but also ensures the preservation of vital anatomic structural content. This approach may improve equity in access to healthcare by enhancing portable device capabilities, although further validation and optimizations are necessary for broader clinical application.

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.