CycleULM: A unified label-free deep learning framework for ultrasound localisation microscopy
Authors: Su Yan, Clara Rodrigo Gonzalez, Vincent C. H. Leung, Herman Verinaz-Jadan, Jiakang Chen, Matthieu Toulemonde, Kai Riemer, Jipeng Yan, Clotilde Vié, Qingyuan Tan, Peter D. Weinberg, Pier Luigi Dragotti, Kevin G. Murphy, Meng-Xing Tang
Organizations: Department of Bioengineering, Faculty of Engineering, Imperial College London, London, UK. · Department of Electrical and Electronic Engineering, Faculty of Engineering, Imperial College London, London, UK. · Faculty of Electrical and Computer Engineering, Escuela Superior Politécnica del Litoral (ESPOL), Guayaquil, Ecuador. · Department of Metabolism, Digestion and Reproduction, Faculty of Medicine, Imperial College London, London, UK.
Super-resolution ultrasound via microbubble (MB) localisation and tracking, also known as ultrasound localisation microscopy (ULM), can resolve microvasculature beyond the acoustic diffraction limit. However, significant challenges remain in localisation performance and data acquisition and processing time. Deep learning methods for ULM have shown promise to address these challenges, however, they remain limited by in vivo label scarcity and the simulation-to-reality domain gap. We present CycleULM, the first unified label-free deep learning framework for ULM. CycleULM learns a physics-emulating translation between the real contrast-enhanced ultrasound (CEUS) data domain and a simplified MB-only domain, leveraging the power of CycleGAN without requiring paired ground truth data. With this translation, CycleULM removes dependence on high-fidelity simulators or labelled data, and makes MB localisation and tracking substantially easier. Deployed as modular plug-and-play components within existing pipelines or as an end-to-end processing framework, CycleULM delivers substantial performance gains across both in silico and in vivo datasets. Specifically, CycleULM improves image contrast (contrast-to-noise ratio) by up to 15.3 dB and sharpens CEUS resolution with a 2.5{\times} reduction in the full width at half maximum of the point spread function. CycleULM also improves MB localisation performance, with up to +40% recall, +46% precision, and a -14.0 μm mean localisation error, yielding more faithful vascular reconstructions. Importantly, CycleULM achieves real-time processing throughput at 18.3 frames per second with order-of-magnitude speed-ups (up to ~14.5{\times}). By combining label-free learning, performance enhancement, and computational efficiency, CycleULM provides a practical pathway toward robust, real-time ULM and accelerates its translation to clinical applications.
Figures & tables
Figure 1: Overview of the CycleULM framework. a , CycleULM learns a bidirectional physics-emulating translation between the CEUS frames and a simplified simulated MB-only domain via self-supervised cycle-consistent adversarial training. The translation by MB-DT effectively bridges the domain gap between simulated data and real acquisitions, enabling the downstream processing or model training within a simplified, well-controlled domain without requiring manual annotations or complex simulators. b , The variant CycleULM methods. The modular design of CycleULM offers the flexibility that three networks can be integrated as plug-and-play modules within conventional ULM pipelines (CycleULM-NCC, CycleULM-Decon and CycleULM-Loc), or combined into an end-to-end DL pipeline for fully automated ULM post-processing (CycleULM-E2E). c , MBL-Net takes the MB-only image translated by MB-DT as input and produces four complementary outputs, including a probability map denoting the likelihood of a MB at each pixel, an intensity map estimating the bubble intensities, two sub-pixel coordinate maps that refine the accuracy of MB locations, and three uncertainty maps that quantify confidence in these predictions. d , MBT-Net takes a short sequence of 8 consecutive MB-only frames translated by the MB-DT as input and outputs three maps, including a trajectory-probability map indicating the likelihood of a MB path present at every pixel, a velocity map showing the horizontal velocity components, and a velocity map showing the vertical velocity components, respectively.
Figure 2: CycleULM learns a physics-emulating translation from the CEUS domain into a simplified MB-only domain. Demonstrations on the in silico ULTRA-SR Challenge data in a and the in vivo rabbit kidney data in b show that, by leveraging the temporal consistency from three consecutive CEUS frames, the MB-DT can effectively distinguish MB signals from background clutter. This physics-emulating translation bridges the domain gap between simulated data and real acquisitions, enabling the downstream processing or model training within a simplified, well-controlled domain without requiring manual annotations or complex simulators.
Figure 3: CycleULM improves image contrast and resolution on in silico ULTRA-SR Challenge dataset. a , An example CEUS frame and b , the corresponding MB-only frame translated by the MB-DT. In the MB-only frame, background clutter is significantly suppressed and the PSF size is visibly reduced, resulting in the improved separation of overlapping MBs (yellow arrows). c , MIPs of the raw CEUS dataset and d , the MB-DT output sequence. The MIP after MB-DT preserves the vascular pattern in the raw MIP but exhibits markedly lower background, higher contrast, and finer structural detail. e,f,g , Zoomed-in MIPs and the accumulated GT MB image show a substantial CNR improvement of 9.2 dB after MB-DT. h , Cross-sectional analyses across a single MB and i , two adjacent vessels demonstrate enhanced resolution. The measured FWHM decreased from 355 µm to 178 µm (a 2-fold resolution improvement), and overlapping MB signals became clearly separable after MB-DT.
Figure 4: CycleULM improves image contrast and resolution in an in vivo rabbit kidney dataset. a , An example CEUS frame and b , the corresponding MB-only frame translated by the MB-DT. The MB-only shows markedly reduced background clutter. c , MIPs of the raw CEUS dataset and d , the sequence output by MB-DT. The MIP after MB-DT preserved the vascular architecture in the raw MIP while achieving substantially higher contrast and resolution. e,f , Zoomed-in MIPs show an extraordinary CNR improvement of 15.3 dB after MB-DT g , Cross-sectional analyses across a single MB and h , several adjacent vessels. The measured FWHM decreased from 725 µm to 294 µm after MB-DT, corresponding to a 2.5-fold resolution improvement. Adjacent vascular structures also appear more clearly separated in the MB-DT–translated MIP.
Figure 5: CycleULM significantly improves MB localisation performance on in silico ULTRA-SR Challenge dataset. a , Precision–recall curves. b , Mean localisation error versus recall curves. c , Accumulated GT MB image. d–h , Accumulated super-localised MB maps generated by different methods, each with the same number of detected MBs. i,j , Zoomed-in comparisons of two selected regions from the super-localised MB maps, highlighting differences in localisation accuracy and background suppression.
Table 1: Quantitative evaluation of MB localisation performance on the in silico ULTRA-SR dataset. The results show that our CycleULM methods with MB-DT consistently produced maps with lower MSE and higher SSIM, as well as improved recall, precision, and F1 score compared with baseline methods. Among all methods, CycleULM-Loc achieved the best overall performance (bold), followed by CycleULM-Decon (underlined), demonstrating the effectiveness of CycleULM in improving MB localisation performance.
Figure 6: CycleULM enhances the SR ULM images on the in silico ULTRA-SR Challenge dataset. a , The GT vascular map. b,c , The SR maps generated by the two baseline methods. d-g , The SR maps generated by the CycleULM methods. The CycleULM methods recovered more detailed and continuous microvascular structures compared to both baselines. In particular, CycleULM-E2E reconstructed visibly finer and more complete vessel networks (green arrows). Quantitative evaluation using RMSE supports that all CycleULM methods achieved lower RMSEs than Baseline 1 and Baseline 2, with CycleULM-E2E having the lowest RMSE of 0.2750, indicating the most accurate reconstruction of the vascular map relative to the GT.
Figure 7: CycleULM enhances the SR ULM images on the in vivo rabbit kidney dataset. a , Power Doppler image created from 500 frames. b,c , The SR maps generated by the two baseline methods. d-g , The SR maps generated by the CycleULM methods. Compared with both baselines, the CycleULM methods recovered finer vascular branches and achieved more complete vessel reconstruction, particularly in the renal hilum areas.
Figure 8: CycleULM enhances the SR ULM images in an in vivo rat brain dataset. a,b , The SR maps generated by the two baseline methods. c-f , The SR maps generated by the CycleULM methods. Compared with both baselines, the CycleULM methods recovered finer and more complete cerebral vasculature reconstruction, highlighted by coloured boxes and arrows.
Table 2: CycleULM speeds up the ULM processing to real-time. Processing times by different methods in seconds and their throughputs in frames per second. CycleULM methods are all faster than the two baseline methods. Notably, the end-to-end deep learning model, CycleULM-E2E (underlined), achieved the fastest SR image generation, requiring only 27.2 seconds for 500 frames, reaching a processing throughput of 18.4 frames per second. Each frame had a spatial size of 511 × 659 pixels and was processed with an upsampling factor of 4. Even without speed-specific optimisations, these results demonstrate the strong potential of CycleULM for real-time ULM imaging.
Figure 9: CycleULM enhances the SR ULM images on the in vivo rabbit kidney dataset. a , Power Doppler image created from 500 frames. b , The SR map generated by Baseline 2 method. c,d , The SR maps generated by CycleULM-Decon and CycleULM-E2E methods. Compared with both baselines, the CycleULM methods recovered finer vascular branches and achieved more complete vessel reconstruction, particularly in the renal hilum areas.
Figure 10: Architectures of the Cycle-GAN and training losses for the MB-DT. a,b , Network architectures of generators GAB (MB-DT) and GBA , respectively. c,d , Network architectures of discriminators DA and DB , respectively. e , Structure of residual network (ResNet) block used within the generators to enhance feature representation in the deeper layers. f , Overall training strategy and associated loss functions for learning the physics-emulating domain translation.
Figure 11: Architecture and training losses of the MBL-Net. a , Network architecture of the MBL-Net, including the structure of the residual (ResNet) block used within the network. b , Overall training strategy and corresponding loss functions employed for learning MB localisation in the MB-only domain.
Figure 12: Architecture and training losses of the MBT-Net. a , Network architecture of the MBT-Net, highlighting the ConvLSTM block used to capture temporal dependencies across MB-only frames. b , Overall training strategy and corresponding loss functions employed for learning MB tracking in the MB-only domain.
Figure S1 : Traditional ULM post-processing pipeline.
Figure S2 : Evaluation of the 3-frame input design for MB-DT. a , Power Doppler image created from 500 frames. b , SR map generated by the Baseline 2 method. c,d , SR maps generated by CycleULM-Decon when MB-DT is trained with three-frame input and single-frame input, respectively. Using three consecutive frames, CycleULM-Decon recovers finer vascular branches and achieves more complete vessel reconstruction.
Ultrasound imaging uses sound waves to visualize anatomy and function, with AI improving image reconstruction, diagnosis, measurement, and scan guidance.
Ultrasound is one of the most widely used medical imaging modalities, and recent large vision-language models(VLMs) have shown increasing capabilities in ultrasound image understanding. However, these models fail to provide pixel-level visual evidence aligned with their semantic predictions, and their fine-grained grounding capability in ultrasound remains largely unclear. We introduce UltraG-Bench, a large-scale multi-task benchmark for evaluating pixel-level evidence grounding in ultrasound. UltraG-Bench is built by annotating 40 public ultrasound segmentation datasets spanning 13 anatomical categories, and comprises three progressive tasks: instruction-guided segmentation, evidence-grounded VQA, and evidence-grounded report generation, with 331125, 666779, and 138832 annotations, respectively. Comprehensive evaluation of 14 state-of-the-art models reveals a substantial gap between semantic understanding and fine-grained pixel-level localization. We further propose UltraG-Agent, which combines the semantic reasoning capabilities of a VLM with the ultrasound-specific segmentation capability of UltraSAM3. Experiments show that UltraG-Agent substantially improves both semantic prediction and pixel-level visual grounding. Our dataset and code are available at https://github.com/zhuqh19/UltraG-Bench.
Large vision-language models (LVLMs) have achieved strong performance across many medical imaging tasks, yet their application to ultrasound remains limited due to its inherent complexity and variability. In this work, we revisit what is truly needed to enable real-world ultrasound understanding. Instead of introducing complex architectures or elaborate training strategies, we show that data scale and clinically faithful data alignment are the key factors. We construct a large-scale dataset of 1.5M real-world ultrasound examinations, containing 17.7M images, multi-organ coverage, and paired uncurated clinical reports. Crucially, we organize the data at the examination level, aligning multiple images with their corresponding reports to reflect real clinical workflows. We then fine-tune a standard LVLM using low-rank adaptation (LoRA) on this dataset without task-specific modifications. Surprisingly, this simple recipe already leads to strong performance across diverse ultrasound understanding tasks, outperforming prior methods designed with more complex pipelines. Beyond these results, we present model and data scaling analyses that provide insights into the role of scale in ultrasound LVLMs.
Bingcong Yan, Chunlei Li, Jingliang Hu +3
1MedAI Technology (Wuxi) Co. Ltd. · 2Technical University of Munich
Ultrasound images vary widely across scanners, operators, and anatomical targets, which often causes models trained in one setting to generalize poorly to new hospitals and clinical conditions. The Foundation Model Challenge for Ultrasound Image Analysis (FMC-UIA) reflects this difficulty by requiring a single model to handle multiple tasks, including segmentation, detection, classification, and landmark regression across diverse organs and datasets. We propose a unified multi-task framework based on a transformer visual encoder from the Qwen3-VL family. Intermediate token features are projected into spatial feature maps and fused using a lightweight multi-scale feature pyramid, enabling both pixel-level predictions and global reasoning within a shared representation. Each task is handled by a small task-specific prediction head, while training uses task-aware sampling and selective loss balancing to manage heterogeneous supervision and reduce task imbalance. Our method is designed to be simple to optimize and adaptable across a wide range of ultrasound analysis tasks. The performance improved from 67% to 85% on the validation set and achieved an average score of 81.84% on the official test set across all tasks. The code is publicly available at: https://github.com/saitejalekkala33/FMCUIA-ISBI.git
Ufaq Khan, L. D. M. S. Sai Teja, Ayuba Shakiru +4
MBZUAI, Abu Dhabi, UAE · NIT Silchar, India · Birmingham City University, UK