IEEE Transactions on Medical Imaging

Three-dimensional photoacoustic imaging (3D PAI) commonly relies on sparse sensor arrays, which restrict angular sampling, detection aperture, and instantaneous field-of-view (FOV). Moving the sensor array relative to the target provides an effective route to multi-view imaging and large volume photoacoustic mapping, but accurate fusion of multiple poses conventionally depends on motor feedback or external tracking hardware. Such tracking increases system complexity and can suffer from calibration errors, backlash, and motion instability. Here we introduce PA-SfM, a tracker-free differentiable acoustic structure-from-motion (SfM) framework that recovers sensor array poses directly from photoacoustic measurements. By integrating a differentiable acoustic radiation model with hierarchical optimization and rigid-array constraints, PA-SfM jointly estimates inter-view transformations and reconstructs 3D photoacoustic volumes without external pose measurements. We validated PA-SfM using numerical simulations, in vivo rat kidney and liver imaging with known relative geometry, and a mechanically scanned 3D PAI system. In mechanically rotated mouse liver imaging, PA-SfM produced sharper and more continuous vascular reconstructions than encoder-based fusion. In translational multi-pose imaging, PA-SfM supported expanded FOV vascular mapping without translation-stage pose input. In controlled quantitative validations, PA-SfM achieved high reconstruction fidelity, with PSNRs of 38.90-41.42 dB and SSIMs of 0.9637-0.9864 relative to groundtruth or known-pose reference reconstructions. These results establish PA-SfM as a robust computational framework for tracker-free multi-view and expanded FOV 3D PAI, providing a complete algorithmic foundation for freehand 3D PAI. The source code is publicly available at https://github.com/JaegerCQ/PA-SfM.

Singular value decomposition (SVD) is the reference clutter rejection strategy for medical ultrafast ultrasound imaging, yet SVD filtered images retain a residual noise floor that obscures microvascular signals at depth. Supervised denoising cannot address this gap because clean references do not exist, and most physics-based alternatives require radio-frequency or in-phase/quadrature data that clinical scanners do not expose. We introduce SVD SN2N, a self supervised Noise2Noise framework that operates entirely on the post SVD image and requires neither clean targets nor channel level access. Our central contribution is to show that the residual fluctuations of SVD filtered images are dominated by multiplicative speckle, which violates the zero mean additive assumption of Noise2Noise theory, and that a single log compression step converts it into an additive perturbation whose variance no longer depends on the signal, restoring the conditions the framework requires. The log compressed image is split into two statistically independent half size copies by diagonal 2 times 2 resampling, rescaled by Fourier zero padding interpolation, and used to train a U Net under a self constrained twin-prediction loss. We assess the underlying assumptions directly by measuring the residual noise distribution, its spatial correlation and the signal similarity within diagonal pixel pairs, and we validate the framework on four heterogeneous datasets spanning preclinical and clinical, contrast enhanced and contrast-free, and 2D to 3D regimes. Relative to conventional SVD, SVD SN2N raises image derived SNR by 4.2 to10.6 dB and narrows the apparent vessel full width at half maximum by 34% to 59%, providing a post SVD denoising front end compatible with the image only data available on clinical scanners.

Textures and edges contribute different information to image recognition. Edges and boundaries encode shape information, while textures manifest the appearance of regions. Despite the success of Convolutional Neural Networks (CNNs) in computer vision and medical image analysis applications, predominantly only texture abstractions are learned, which often leads to imprecise boundary delineations. In medical imaging, expert manual segmentation often relies on organ boundaries; for example, to manually segment a liver, a medical practitioner usually identifies edges first and subsequently fills in the segmentation mask. Motivated by these observations, we propose a plug-and-play module, dubbed Edge-Gated CNNs (EG-CNNs), that can be used with existing encoder-decoder architectures to process both edge and texture information. The EG-CNN learns to emphasize the edges in the encoder, to predict crisp boundaries by an auxiliary edge supervision, and to fuse its output with the original CNN output. We evaluate the effectiveness of the EG-CNN with various mainstream CNNs on two publicly available datasets, BraTS 19 and KiTS 19 for brain tumor and kidney semantic segmentation. We demonstrate how the addition of EG-CNN consistently improves segmentation accuracy and generalization performance.

Lesion segmentation is an important problem in computer-assisted diagnosis that remains challenging due to the prevalence of low contrast, irregular boundaries that are unamenable to shape priors. We introduce Deep Active Lesion Segmentation (DALS), a fully automated segmentation framework that leverages the powerful nonlinear feature extraction abilities of fully Convolutional Neural Networks (CNNs) and the precise boundary delineation abilities of Active Contour Models (ACMs). Our DALS framework benefits from an improved level-set ACM formulation with a per-pixel-parameterized energy functional and a novel multiscale encoder-decoder CNN that learns an initialization probability map along with parameter maps for the ACM. We evaluate our lesion segmentation model on a new Multiorgan Lesion Segmentation (MLS) dataset that contains images of various organs, including brain, liver, and lung, across different imaging modalities--MR and CT. Our results demonstrate favorable performance compared to competing methods, especially for small training datasets.

Automated segmentation of mediastinal neoplasms in preoperative computed tomography (CT) scans is critical for accurate diagnosis. Though convolutional neural networks (CNNs) have proven effective in medical imaging analysis, the segmentation of mediastinal neoplasms, which vary greatly in shape, size, and texture, presents a unique challenge due to the inherent local focus of convolution operations. To address this limitation, we propose a confidence-enhanced semi-supervised learning framework for mediastinal neoplasm segmentation. Specifically, we introduce a confidence-enhanced module that improves segmentation accuracy over indistinct tumor boundaries by assessing and excluding unreliable predictions simultaneously, which can greatly enhance the efficiency of exploiting unlabeled data. Additionally, we implement an iterative learning strategy designed to continuously refine the estimates of prediction reliability throughout the training process, ensuring more precise confidence assessments. Quantitative analysis on a real-world dataset demonstrates that our model significantly improves the performance by leveraging unlabeled data, surpassing existing semi-supervised segmentation benchmarks. Finally, to promote more efficient academic communication, the analysis code is publicly available at https://github.com/fxiaotong432/CEDS. Author summaryIn clinical practice, computed tomography (CT) scans can aid in the detection and evaluation of mediastinal tumors. The early detection of mediastinal tumors plays a crucial role in formulating appropriate treatment plans and improving patient survival rates. To reduce the high cost of manual annotation, researchers have attempted to employ convolutional neural networks (CNNs) for efficient automatic segmentation. However, the significant challenges arise due to the considerable variation in shape, size, and texture of mediastinal tumors, posing difficulties for the segmentation task. In this study, we introduce a confidence-enhanced module with a semi-supervised learning framework. By evaluating the models prediction confidence and selecting high-confidence predictions, we improve the efficiency and quality of data utilization. This approach demonstrates the achievement of accurate mediastinal tumor segmentation with only a minimal amount of labeled data. Our research not only provides an effective technical approach for automatic segmentation of mediastinal tumors but also opens up new possibilities for optimizing strategies in semi-supervised learning methods.

Voxel-level annotation for volumetric medical imaging is expensive and difficult to scale, which makes training highcapacity 3-D segmentation models challenging in practice. Transfer learning (TL) from large public datasets is a common remedy, but it can under-perform when the source domain differs from the target anatomy and acquisition characteristics, as is often the case for pulmonary nodules. In this work, we propose a masked autoencoder (MAE) pretraining-based approach to break the data efficiency wall of domain difference and present a focused empirical study of domain-specific self-supervised learning (SSL) for 3-D lung nodule segmentation. We evaluate two experimental settings: first, Masked Autoencoder (MAE) pretraining versus random initialization across representative baselines; second, MAE versus Decathlon TL for UNETR++ while testing whether MAE-based pretraining also benefits a CNN baseline (V-Net). MAE pretraining on target-domain CT volumes achieves a Dice Similarity Coefficient (DSC) of 0.307, outperforming random initialization (0.136) and Decathlon weights (0.257). In addition, MAE improves the stability of V-Net in a "low-data" regime (i.e., with "insufficiently labeled" data), increasing DSC from 0.010 to0.071. Overall, these results suggest that MAE-based pretraining can provide a practical and robust initialization strategy for volumetric segmentation when labeled data are limited.

Fully convolutional neural networks (CNNs) have proven to be effective at representing and classifying textural information, thus transforming image intensity into output class masks that achieve semantic image segmentation. In medical image analysis, however, expert manual segmentation often relies on the boundaries of anatomical structures of interest. We propose boundary aware CNNs for medical image segmentation. Our networks are designed to account for organ boundary information, both by providing a special network edge branch and edge-aware loss terms, and they are trainable end-to-end. We validate their effectiveness on the task of brain tumor segmentation using the BraTS 2018 dataset. Our experiments reveal that our approach yields more accurate segmentation results, which makes it promising for more extensive application to medical image segmentation.

ObjectiveSuper-resolution ultrasound (SR-US) reveals microvascular structures with exquisite resolution, but clinical translation remains limited by the need for ultrafast frame rates, massive data volumes, and long reconstruction times. This work proposes a deep learning framework that reconstructs microvascular maps from low-frame-rate enhanced ultrasound sequences, bypassing explicit microbubble localization and tracking. MethodsA transformer-decoder network with learned linear projections was designed to model spatiotemporal dependencies across sparse contrast-enhanced ultrasound sequences and reconstruct vessel probability maps, refined via a post-processing enhancement stage. Single-head self-attention captures temporal correlations under challenging conditions including overlapping microbubbles and low signal-to-noise ratios. Binary cross-entropy loss guided training to preserve vascular topology across synthetic and in vivo datasets. In vivo rat brain bolus data from the PALA challenge was used to evaluate this approach under up to 500 - fold data reduction (341 frames at 2 FPS vs. 170400 frames at 1000 FPS in standard ULM). ResultsDespite aggressive undersampling, the proposed pipeline recovered coherent microvascular architecture where conventional ULM pipelines applied to the same sparse data failed to produce continuous vascular networks. Major branches and higher-order microvessels remained visible with apparent vessel widths broadened by approximately three-fold relative to reference SR-US. End-to-end reconstruction completed in 28-133 seconds on an NVIDIA H100 GPU depending on the number of frames employed. ConclusionThe reported approach preserved vascular topology with fast reconstruction and low data overhead, albeit at lower resolution. The substantial reduction in frames and computation time highlights the translational potential of this SR-US-inspired microvascular imaging approach.

Reduced angular sampling is a key strategy for increasing scanning efficiency of micron-scale computed tomography (micro-CT). Despite boosting throughput, this strategy introduces noise and artifacts due to undersampling. In this work, we present a solution to this issue, by proposing a novel Dense Residual Hierarchical Transformer (DRHT) network to recover high-quality sinograms from 2 x, 4x and 8x undersampled scans. DRHT is trained to utilize limited information available from sparsely angular sampled scans and once trained, it can be applied to recover higher-resolution sinograms from shorter scan sessions. Our proposed DRHT model aggregates the benefits of a hierarchical-multi-scale structure along with the combination of local and global feature extraction through dense residual convolutional blocks and non-overlapping window transformer blocks respectively. We also propose a novel noise-aware loss function named KL-L1 to improve sinogram restoration to full resolution. KL-L1, a weighted combination of pixel-level and distribution-level cost functions, leverages inconsistencies in noise distribution and uses learnable spatial weights to improve the training of the DRHT model. We present ablation studies and evaluations of our method against other state-of-the-art (SOTA) models over multiple datasets. Our proposed DRHT network achieves an average increase in peak signal to noise ratio (PSNR) of 17.73dB and a structural similarity index (SSIM) of 0.161, for 8x upsampling, across the three unique datasets, compared to their respective Bicubic interpolated versions. This novel approach can be utilized to decrease radiation exposure to patients and reduce imaging time for large-scale CT imaging projects.

Magnetic resonance imaging (MRI) exhibits substantial scanner and protocol dependent appearance variations, making harmonization challenging without distorting patient specific anatomy, particularly in abdominal imaging. We propose AC-Diff, an anatomy--contrast disentangled diffusion framework that formulates MRI harmonization as a factor-specific generative intervention. Using aligned multi-contrast supervision and cross-patient factor swapping, AC-Diff learns to separate anatomical structure from acquisition-dependent contrast. Unlike conventional image or latent diffusion, AC-Diff restricts stochastic generation to the contrast subspace while the source anatomy bypasses diffusion and is directly reused during reconstruction. Experiments on an in-house multi-vendor cohort and the external Duke Liver MRI dataset demonstrate improved target-domain alignment with strong structural preservation. In downstream liver segmentation, AC-Diff improves Dice from 0.942 for the original inputs to 0.954 for the harmonized images. These results support contrast-specific latent generation as a promising approach to anatomy-preserving MRI harmonization.

Fetal brain magnetic resonance imaging is of great importance for prenatal neural disorder diagnosis. To improve signal-to-noise ratio and coverage, fetal brain MRI often uses thick-slice scanning to reduce motion artifacts and ensure image quality for single slices. Reconstructing 3D MR volumes from multiple motion-corrupted stacks of 2D slices has shown promise in imaging of moving subjects like fetal MRI. However, raw scans acquired with varying slice thicknesses present substantial disparities in quality when retrospectively reconstructed into isotropic high-resolution volumes (e.g. 0.8 mm slice thickness). In particular, thick-slice acquisitions (e.g. 5-6 mm) tend to yield suboptimal reconstruction results compared to thin-slice scans (e.g. 2-3 mm), often exhibiting residual noise, interpolation-induced artifacts, and inter-slice structural discontinuities. This challenge has highlighted the need to further reduce noise caused by reconstruction errors when the reconstructed volumes are used for downstream tasks. With the success of neural networks in computer vision, deep learning methods based on architectures such as convolutional neural networks (CNNs) and generative adversarial networks (GANs) have achieved outstanding performance in single-image denoising. Nevertheless, CNN-based methods still heavily rely on large-scale datasets comprising noisy-clean image pairs. Although unsupervised GAN variants such as CycleGAN have been developed to mitigate this dependency, the inherent variability of GAN-generated outputs often results in tiny anatomical distortions, significantly limiting the applicability of GAN-based methods. While several approaches have introduced regularizations to address this issue, they largely focus on denoising individual slices, overlooking inter-slice structural inconsistencies that arise from treating slices independently. In this work, we propose Pseudo-Warping Field Guided Unsupervised Generative Adversarial Network (PW-GAN), which formulates post-reconstruction optimization as an unpaired style transfer problem between low-quality and high-quality MRI domains. Moreover, by incorporating a pseudo-deformation field module based on optical flow estimation, our method significantly enhances inter-slice continuity in the reconstructed volumes while effectively suppressing residual noise and interpolation artifacts introduced during the reconstruction process. Evaluations on both simulated and in vivo data demonstrate that our method outperforms existing unsupervised models and achieves performance on par with several state-of-the-art supervised methods.

In this paper, we present a novel and efficient framework for cross-modality medical image synthesis, developed for BraSyn-Task 8. Our method combines the fast-sampling capabilities of the Fast-Denoising Diffusion Probabilistic Model (Fast-DDPM) with Discrete Wavelet-Transformed components, as used in Conditional Wavelet Diffusion Models. By reducing the number of denoising steps to 100 and using wavelet-transformed inputs, we accelerate both training and inference and reduce memory usage while preserving high image quality. The framework was trained on the BraTS 2025 dataset, which includes four magnetic resonance imaging (MRI) modalities: T1-weighted, contrast-enhanced T1-weighted (T1c), T2-weighted, and FLAIR. We developed four independent models, each synthesizing one missing modality from the remaining three. Evaluation on the BraSyn 2025 Task 8 public validation set demonstrated competitive performance using standard image metrics: mean squared error, signal-to-noise ratio, and structural similarity index. Our method achieved Third place in the challenge in the final test data, with fast inference times (average 41- 67 seconds per case). To assess clinical relevance, we applied a pretrained nnU-Net segmentation model on the synthesized modalities. Segmentation results yielded high Dice coefficients: 0.877 for the whole tumor, 0.769 for the tumor core, and 0.667 for the enhancing tumor. These results confirm the effectiveness and reliability of our approach for missing-modality synthesis, enabling accurate downstream analysis in high-dimensional medical imaging tasks. Our team in the challenge is USD-2025-Chato-Sereda (Team ID: 3551654).Github link: https://github.com/tsereda/brats-synthesis

Much of the information needed for diagnosis and treatment monitoring of diseases like cancer and cardiovascular disease is found at scales below the resolution limit of classic ultrasound imaging. Recently introduced vascular super-localization methods provide more than a ten-fold improvement in spatial resolution by precisely estimating the positions of microbubble contrast agents. However, most vascular ultrasound scans are currently performed without contrast agents due to the associated cost, training, and post-scan monitoring. Here we show that super-resolution ultrasound imaging of dense vascular structures can be achieved using the natural contrast of flowing blood cells. Instead of relying on separable targets, we used Fourier-based decomposition to separate signals arising from the different scales of vascular structures while removing speckle noise using multi-ensemble processing. This approach enabled the use of compressed sensing for super-resolution imaging of the underlying vascular structures, improving resolution by a factor of four. Reconstruction of ultrafast mouse brain scans revealed details that could not be resolved in regular Doppler images, agreeing closely with bubble-based super-localization microscopy of the same fields of view. By combining multi-ensemble Doppler acquisitions with narrowband Fourier decomposition and computational super-resolution imaging, this approach opens new opportunities for affordable and scalable super-resolution ultrasound imaging.

The cerebral vasculature is central to brain function, with alterations linked to numerous cerebrovascular and neurological disorders. Yet, no single imaging modality can capture the entire cerebral vascular network in humans. Instead, an array of techniques are sensitized to different spatial scales, while trading off resolution for coverage. Magnetic Resonance Imaging (MRI) typically resolves only large pial vessels, while high-resolution microscopy allows micrometer-scale vessels to be mapped over limited spatial extents. These techniques must therefore be combined to obtain a complete mapping of the cerebral angioarchitecture, which underscores the need for automatic, cross-modal vessel segmentation. Here, we introduce VesSynth, a flexible vessel segmentation framework that achieves state-of-the-art accuracy across multiple modalities and spatial resolutions (MR, optical and X-ray imaging), despite being trained entirely on synthetic data. By enabling consistent vascular mapping across scales, this framework paves the way to comprehensive investigation of cerebrovascular organization and its role in health and disease.

Noncontact dynamic optical coherence elastography (OCE) can measure the elastic properties of anterior soft tissues. However, applying OCE to skin requires tracking mechanical waves propagating in multiple directions to reconstruct mechanical anisotropy. Robotic scanning is well suited to this task but requires accurate probe positioning, alignment to the surface normal, and precise, repeatable rotation. To ensure phase and polarization stability, polarization-maintaining (PM) fibers can be used, although they introduce stripe and ghost artifacts that can destabilize surface detection. We present DINOCT, a DINO-based method to localize the skin surface from optical coherence tomography (OCT) images that combines self-supervised pretraining with DINO (self-DIstillation with NO labels) and iBOT (image BERT pre-training with Online Tokenizer) objectives, low-rank adaptation, and a lightweight curve decoder. Compared with classical and supervised learning baselines, DINOCT achieved competitive localization accuracy on clean test data with a near-zero spike rate. Among the learned methods, DINOCT achieved the lowest average and maximum values of mean absolute error (MAE) for severe synthetic stress conditions and the lowest catastrophic failure rates on a separate recording with strong PM-fiber artifacts.

Ultrafast power Doppler imaging combined with singular value decomposition (SVD) clutter filtering has become a standard approach for label-free microvascular ultrasound, enabling the visualization of small vessels without microbubble contrast agents. In the absence of contrast, however, SVD filtered Doppler images remain limited by the blur of the imaging system point spread function (PSF) and by a residual noise floor that reduces sensitivity at depth, which together hinder the resolution of fine microvasculature. Here we establish a sparse deconvolution framework to SVD filtered ultrafast power Doppler images. Each Doppler frame is processed in two cascaded stages: a Split-Bregman optimization that solves a regularized least-squares problem combining a sparsity prior and a Hessian continuity prior, followed by an accelerated Richardson Lucy deconvolution with an estimated system PSF. We first validated the framework on a simulation phantom with known ground truth, and then evaluated the framework on in vivo rat-brain plane-wave acquisitions obtained with a Verasonics Vantage system and a 15-MHz linear array. Compared to conventional SVD power Doppler, sparse deconvolution improved the resolution by around 4 and 8 times to lambda/2 and lambda/4, in the lateral and axial directions respectively. We further show that decomposing the Doppler signal into velocity bands before deconvolution disentangles slow and fast flow and yields a velocity-resolved microvascular map. Finally, applying the same framework to task-evoked functional ultrasound, we show that sparse deconvolution preserves the stimulus-locked cerebral-blood-volume response measured by conventional functional ultrasound while sharpening the corresponding activation map from a diffuse cortical region to discrete penetrating vessels. These results indicate that the sparsity-prior super-resolution principles established in label-free ultrafast Doppler ultrasound, and that sparse deconvolution can serve as a practical, contrast-free post-processing front-end for super-resolution microvascular and functional imaging.

Three-dimensional photoacoustic tomography (3D-PAT) enables noninvasive structural and functional imaging with optical absorption contrast and ultrasonic detection depth. However, its spatial resolution is limited by acoustic diffraction, and incomplete detection geometry can substantially degrade image fidelity and quantitative accuracy. Here, we present a ULM-guided model-based reconstruction framework, termed 3D-PAULMprior that incorporates sub-diffraction vascular priors from concurrent ultrasound localization microscopy (ULM) into 3D photoacoustic reconstruction. The method uses weighted regional Laplacian regularization to integrate high-resolution vascular information into the inverse problem, thereby enhancing vascular sharpness, suppressing limited-view artifacts, and improving blood oxygen saturation estimation. We validated 3D-PAULMprior using numerical simulations, tissue-mimicking phantoms, and in vivo mouse brain imaging. Compared with conventional reconstruction, 3D- PAULMprior improved spatial resolution by over 50%, increased contrast-to-noise ratio by 261.2%, and enhanced structural similarity index by 24.6%. In vivo, 3D-PAULMprior recovered vascular structures that were poorly resolved or missing in conventional reconstructions and produced more spatially confined sO2 maps. These results establish 3D-PAULMprior as a robust multimodal reconstruction strategy for high-resolution structural and functional photoacoustic imaging.

Ultrasound localization microscopy (ULM) enabled high-accuracy measurements of microvessel flow beyond the resolution limit of conventional ultrasound imaging by utilizing contrast microbubbles (MBs) as point targets. Robust tracking of MBs is an essential task for fast and high-quality ULM image reconstruction. Existing MB tracking methods suffer from challenging imaging scenarios such as high-density MB distributions, fast blood flow, and complex flow dynamics. Here we present a deep learning-based MB pairing and tracking method based on a bi-directional long short-term memory neural network for ULM. The proposed method integrates multiparametric MB characteristics to facilitate more robust and accurate MB pairing and tracking. The method was validated on a simulation data set, a tissue-mimicking flow phantom, and in vivo on a mouse and rat brain.

Clearing Assisted Scattering Tomography (CAST) extends coherent scattering tomography to whole-brain imaging, enabling visualization of fine-scale brain-wide connectivity. As a coherent optical tomography modality, CAST is inherently affected by speckle noise, which degrades image quality and limits quantitative analysis. However, existing speckle reduction methods developed for optical coherence tomography (OCT) are not directly transferable to CAST images due to differences in sample and noise statistics. Here, we present a learning-based cleared-sample speckle reduction network, termed CLEAR Net, specifically designed for CAST imaging, which effectively suppresses speckle noise in whole-brain white matter images while preserving fine structural details. We quantitatively benchmarked CLEAR Net against representative speckle reduction algorithms on CAST datasets and further evaluated its generalizability using publicly available ophthalmic datasets.

Segmenting noisy multiplex spatial tissue images constitutes a challenging task, since the characteristics of both the noise and the biology being imaged differs significantly across tissues and modalities; this is compounded by the high monetary and time costs associated with manual annotations. It is therefore imperative to build algorithms that can accurately segment the noisy images based on a small number of annotations. Recently techniques to derive such an algorithm from a few scribbled annotations have been proposed, mostly relying on the refinement and estimation of pseudo-labels. Other techniques leverage the success of self-supervised denoising as a parallel task to potentially improve the segmentation objective when few annotations are available. In this paper, we propose a method that augments the segmentation objective via self-supervised multi-channel quantized imputation, meaning that each class of the segmentation objective can be characterized by a mixture of distributions. This approach leverages the observation that perfect pixel-wise reconstruction or denoising of the image is not needed for accurate segmentation, and introduces a self-supervised classification objective that better aligns with the overall segmentation goal. We demonstrate the superior performance of our approach for a variety of cancer datasets acquired with different highly-multiplexed imaging modalities in real clinical settings. Code for our method along with a benchmarking dataset is available at https://github.com/natalialmg/ImPartial.