Implicit Neural Representations (INRs) encode signals as the weights of a coordinate-based neural network and have recently been proposed as an alternative domain for downstream learning. While promising, classification directly in weight space remains challenging due to the high dimensionality and complex structure of INR parameters. Furthermore, the way discriminative information is distributed across INR weights remains poorly understood. We propose a hierarchical Mixture-of-Experts (HMoE) Transformer that processes INR weights using conditional computation aligned with the structure of the underlying implicit network. Coupled with a meta-learning framework that shapes INR parameters for downstream tasks, our model achieves state-of-the-art accuracy across standard benchmarks, ranging from low-resolution datasets to high-resolution ImageNet-1K. To gain insight into how INRs encode discriminative information, we develop weight-space attribution and pruning methods that identify parameters most relevant for classification. These analyses reveal how class-specific structure emerges within INR layers and support the suitability of MoE architectures for weight-space learning. Our approach advances both the performance and interpretability of weight-space classifiers.

BACKGROUND:Attention-Deficit/Hyperactivity Disorder (ADHD) is a common neurodevelopmental disorder marked by attentional dysregulation and executive dysfunction. Despite its prevalence, diagnosis remains primarily clinical, lacking robust, objective biomarkers. NEW METHOD:This study introduces a non-invasive, data-driven method for identifying ADHD-related neural signatures using electroencephalography (EEG) combined with Lempel-Ziv Complexity (LZC). LZC quantifies the temporal complexity of neural signals, offering a sensitive measure of brain dynamics that may reveal subtle differences between clinical and control groups. RESULTS:Our spatiotemporal analysis revealed that LZC effectively differentiates ADHD from control EEG patterns at the level of local temporal signal dynamics, with the most pronounced group separations observed at right-lateralized parietal and frontal electrodes, particularly P8, P7, and F8, suggesting altered local EEG signal dynamics over regions implicated in visuospatial attention and executive control. Additional sites, including T7, Pz, O1, and Fp1, exhibited moderate discriminative power, whereas midline and central regions showed minimal differences. Children with ADHD showed reduced nonlinear EEG complexity compared with controls, reflecting altered local temporal EEG signal dynamics during task engagement, particularly at right-lateralised electrodes. The most pronounced group differences were observed at P8 and F8 across both earlier and later segments of the recording, indicating consistently lower signal complexity in the ADHD group. This pattern suggests that right parietal-temporal and frontal regions carry the highest discriminative power for distinguishing children with ADHD from typically developing controls. COMPARISON WITH EXISTING METHODS:Unlike conventional EEG analyses focusing on power spectra or event-related potentials, the LZC-based method captures dynamic complexity, enabling more nuanced differentiation of neural activity patterns. This complexity metric provides complementary information that traditional frequency or time-domain approaches may overlook. CONCLUSIONS:LZC analysis of EEG signals offers a promising complementary biomarker for ADHD, highlighting altered neural dynamics in regions implicated in attentional and executive processes. This complexity-based approach may enhance the objectivity and precision of neurodevelopmental disorder diagnostics.

Vision Transformers (ViTs) achieve strong image-recognition performance, but their parameter count grows linearly with depth when each block is independently parameterized. Single-block recurrent ViTs (bViT) remove this growth by repeatedly applying one shared block. Rather than proposing a new architecture, we fix a bViT and provide a controlled empirical characterization of three training and inference regimes under a common CIFAR-100 protocol, asking: (i) when does recurrence beat independently parameterized depth—at matched FLOPs or at matched parameter memory? (ii) when a residual recurrent block is trained through an ODE solver, does solver order act as numerical refinement or as an architectural bias? and (iii) what does robustness beyond the training horizon cost in nominal accuracy? We find that standard ViTs remain preferable when FLOPs are the primary constraint, whereas recurrent ViTs offer a better accuracy–parameter trade-off under memory constraints. Consistent with the standard view of residual networks as Euler discretizations of ODEs, the continuous-time analogue of a residual recurrent block is the state-subtracted vector field ż=F_θ(z)-z; although known in principle, this distinction is easy to violate when the block is wrapped as a black-box vector field, and we qualify the cost at few accuracy points. Because the vector field is learned jointly with the solver, higher-order solvers act as a solver-induced architectural bias rather than a numerical-accuracy improvement, and their gains are not uniform. Finally, stage-wise deep supervision traces an accuracy–robustness frontier: it does not improve nominal accuracy, but degrades gracefully far beyond the training horizon, where naive recurrence collapses to near-random performance.

We present PatchMorph, a novel stochastic framework designed to significantly reduce memory demand during training and inference of existing unsupervised convolutional and transformer-based registration networks for 3D brain image registration. PatchMorph extends existing methods to handle global registration tasks that are impractical with large image sizes and differing array dimensions, and is also designed to handle varying voxel resolutions, which are common in real-world datasets. It supports the integration of a VoxelMorph-like architecture, CNN or Transformer, and, in experiments, reduced the training memory requirement of a transformer-based network on [Formula: see text] voxel images from around 40 GB to under 10 GB while maintaining performance. PatchMorph decouples spatial logic from network architecture. It operates by matching patches of constant size across multiple scales, from coarse to fine, combining global transformations with local deformations. Unlike conventional multiscale cascade networks, PatchMorph performs geometrically linked, patchwise cascading in world coordinate space. Patch placement and resolution are managed independently of the registration network, enabling the reuse of compact networks with minimal memory footprint. At finer scales, each patch zooms in on a region of a coarser level patch, ensuring spatial hierarchy and continuity. This patch-based, coarse-to-fine refinement in world coordinates is technically challenging, as patches at different scales capture varying image content and must propagate deformation information coherently. PatchMorph addresses these challenges through a modular architecture that preserves spatial consistency across scales and orientations. Experiments on human T1 MRI and marmoset brain images from serial two-photon tomography demonstrate that PatchMorph achieves state-of-the-art registration performance while maintaining a significantly reduced memory footprint. By enabling the use of sophisticated architectures on large, high-resolution, or heterogeneously sampled data, PatchMorph removes a major bottleneck in the development and deployment of deep learning-based image registration networks. While this patch-based approach incurs a higher inference time than single-shot full-volume networks, it remains significantly faster than classical iterative methods.

Ultrasound (US) is a widely used imaging modality due to its availability, relatively low cost, and high frame rate. However, the quality of US images is usually considered inferior compared to other modalities, which can make image interpretation and the development of computer-aided diagnostic systems more challenging. In this work, we propose a deep reinforcement learning approach for US image denoising. Our method employs a software agent that learns to select both the location and types of interpretable image processing filters to apply. We demonstrate that the agent effectively reduces speckle noise in homogeneous regions while preserving or enhancing the structural details at the boundaries between such regions. Compared to other deep learning methods, our approach is based on well-known and simple filtering operations, making the denoising process more transparent and easier to interpret.

Ultrasound Computer Tomography (USCT) enables imaging of the spatial distribution of physical parameters, such as speed of sound and acoustic attenuation. These parameters are highly relevant in the context of tissue diagnostics, as pathological regions, for example tumors, often exhibit distinct attenuation and speed-of-sound characteristics compared to healthy tissue. Achieving USCT imaging at a clinically reasonable frame rate presents a number of challenges. The frame rate is constrained by several factors: the time of flight of ultrasound pulses between the source and the receivers, the number of transmit/receive events, the transfer of acquired data to the host computer, and the time required for data processing and image reconstruction. Among these, the physical limitations imposed by the time of flight-necessary for a pulse to fully traverse the imaged object-are particularly critical. While data transfer and processing times may be reduced through hardware acceleration and software optimization, these fundamental physical constraints cannot be circumvented in the same manner. Consequently, the reduction of the number of transmissions remains the only practical strategy to decrease this component of the total acquisition time. Although such a reduction generally results in some deterioration of image quality, this degradation can be mitigated through interpolation. This strategy shortens both the acquisition and data transfer stages, while shifting the computational burden to more intensive data processing. The objective of the present study is to demonstrate a proof of concept for this approach in USCT imaging, employing a neural network-based interpolator to restore image quality.

Deep learning methods are gaining momentum in radiology. In this work, we investigate the usefulness of vision-language models (VLMs) and large language models for binary few-shot classification of medical images. We utilize the GPT-4 model to generate text descriptors that encapsulate the shape and texture characteristics of objects in medical images. Subsequently, these GPT-4 generated descriptors, alongside VLMs pre-trained on natural images, are employed to classify chest X-rays and breast ultrasound images. Our results indicate that few-shot classification of medical images using VLMs and GPT-4 generated descriptors is a viable approach. However, accurate classification requires the exclusion of certain descriptors from the calculations of the classification scores. Moreover, we assess the ability of VLMs to evaluate shape features in breast mass ultrasound images. This is performed by comparing VLM-based results generated for shape-related text descriptors with the actual values of the shape features calculated using segmentation masks. We further investigate the degree of variability among the sets of text descriptors produced by GPT-4. Our work provides several important insights about the application of VLMs for medical image analysis.

Thyroid nodule assessment is an important problem in ultrasound (US) imaging. In this work, we propose a novel approach to thyroid lesion classification in contrast-enhanced ultrasound (CEUS). We develop implicit neural representations (INRs) that jointly encode B-mode and CEUS sequences, mapping spatio-temporal coordinates to both image types. INRs are coordinate-based neural networks that store information in their weights, enabling continuous and compact representations of medical image data. This representation naturally captures texture and contrast dynamics relevant to nodule characterization, without the need for handcrafted features or specialized recurrent architectures. In our study, each CEUS sequence is encoded using a SIREN INR, and a probing technique is then applied in the weight space of the trained networks to differentiate malignant and benign nodules. We evaluated our method on CEUS data, showing that it achieves good classification performance. To our knowledge, this is the first study to apply INRs to CEUS analysis for thyroid nodule assessment, highlighting their potential for deep learning based tissue characterization in US.

Explaining deep learning models in a way that humans can easily understand is essential for responsible artificial intelligence applications. Attribution methods constitute an important area of explainable deep learning. The attribution problem involves finding parts of the network's input that are the most responsible for the model's output. In this work, we demonstrate that implicit neural representations (INRs) constitute a good framework for generating visual explanations. Firstly, we utilize coordinate-based implicit networks to reformulate and extend the extremal perturbations technique and generate attribution masks. Experimental results confirm the usefulness of our method. For instance, by proper conditioning of the implicit network, we obtain attribution masks that are well-behaved with respect to the imposed area constraints. Secondly, we present an iterative INR-based method that can be used to generate multiple non-overlapping attribution masks for the same image. We depict that a deep learning model may associate the image label with both the appearance of the object of interest as well as with areas and textures usually accompanying the object. Our study demonstrates that implicit networks are well-suited for the generation of attribution masks and can provide interesting insights about the performance of deep learning models.

Monitoring the healing progress of diabetic foot ulcers is a challenging process. Accurate segmentation of foot ulcers can help podiatrists to quantitatively measure the size of wound regions to assist prediction of healing status. The main challenge in this field is the lack of publicly available manual delineation, which can be time consuming and laborious. Recently, methods based on deep learning have shown excellent results in automatic segmentation of medical images, however, they require large-scale datasets for training, and there is limited consensus on which methods perform the best. The 2022 Diabetic Foot Ulcers segmentation challenge was held in conjunction with the 2022 International Conference on Medical Image Computing and Computer Assisted Intervention, which sought to address these issues and stimulate progress in this research domain. A training set of 2000 images exhibiting diabetic foot ulcers was released with corresponding segmentation ground truth masks. Of the 72 (approved) requests from 47 countries, 26 teams used this data to develop fully automated systems to predict the true segmentation masks on a test set of 2000 images, with the corresponding ground truth segmentation masks kept private. Predictions from participating teams were scored and ranked according to their average Dice similarity coefficient of the ground truth masks and prediction masks. The winning team achieved a Dice of 0.7287 for diabetic foot ulcer segmentation. This challenge has now entered a live leaderboard stage where it serves as a challenging benchmark for diabetic foot ulcer segmentation.

Accurate estimation of the speed-of-sound (SoS) is important for ultrasound (US) image reconstruction techniques and tissue characterization. Various approaches have been proposed to calculate SoS, ranging from tomography-inspired algorithms like CUTE to convolutional networks, and more recently, physics-informed optimization frameworks based on differentiable beamforming. In this work, we utilize implicit neural representations (INRs) for SoS estimation in US. INRs are a type of neural network architecture that encodes continuous functions, such as images or physical quantities, through the weights of a network. Implicit networks may overcome the current limitations of SoS estimation techniques, which mainly arise from the use of non-adaptable and oversimplified physical models of tissue. Moreover, convolutional networks for SoS estimation, usually trained using simulated data, often fail when applied to real tissues due to out-of-distribution and data-shift issues. In contrast, implicit networks do not require extensive training datasets since each implicit network is optimized for an individual data case. This adaptability makes them suitable for processing US data collected from varied tissues and across different imaging protocols.We evaluated the proposed SoS estimation method based on INRs using data collected from a tissue-mimicking phantom containing four cylindrical inclusions, with SoS values ranging from 1480 m/s to 1600 m/s. The inclusions were immersed in a material with an SoS value of 1540 m/s. In experiments, the proposed method achieved strong performance, clearly demonstrating the usefulness of implicit networks for quantitative US applications.

In this study, we introduce “instance loss functions”, a new family of loss functions designed to enhance the training of neural networks in the instance-level segmentation and detection of objects in biomedical image data, particularly those of varied numbers and sizes. Intended to be utilized conjointly with traditional loss functions, these proposed functions, prioritize object instances over pixel-by-pixel comparisons. The specific functions, the instance segmentation loss (Linstance), the instance center loss (Lcenter), the false instance rate loss (Lfalse), and the instance proximity loss (Lproximity), serve distinct purposes. Specifically, Linstance improves instance-wise segmentation quality, Lcenter enhances segmentation quality of small instances, Lfalse minimizes the rate of false and missed detections across varied instance sizes, and Lproximity improves detection quality by pulling predicted instances towards the ground truth instances. Through the task of segmenting white matter hyperintensities (WMH) in brain MRI, we benchmarked our proposed instance loss functions, both individually and in combination via an ensemble inference models approach, against traditional pixel-level loss functions. Data were sourced from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and the WMH Segmentation Challenge datasets, which exhibit significant variation in WMH instance sizes. Empirical evaluations demonstrate that combining two instance-level loss functions through ensemble inference models outperforms models using other loss function on both the ADNI and WMH Segmentation Challenge datasets for the segmentation and detection of WMH instances. Further, applying these functions to the segmentation of nuclei in histopathology images demonstrated their effectiveness and generalizability beyond WMH, improving performance even in contexts with less severe instance imbalance.

Pairwise image registration is a necessary prerequisite for brain image comparison and data integration in neuroscience and radiology. In this work, we explore the efficacy of implicit neural representations (INRs) in improving the performance of brain image registration in magnetic resonance imaging. In this setting, INRs serve as a continuous and coordinate based approximation of the deformation field obtained through a multi-layer perceptron. Previous research has demonstrated that sinusoidal representation networks (SIRENs) surpass ReLU models in performance. In this study, we first broaden the range of activation functions to further investigate the registration performance of implicit networks equipped with activation functions that exhibit diverse oscillatory properties. Specifically, in addition to the SIRENs and ReLU, we evaluate activation functions based on snake, sine+, chirp and Morlet wavelet functions. Second, we conduct experiments to relate the hyper-parameters of the models to registration performance. Third, we propose and assess various techniques, including cycle consistency loss, ensembles and cascades of implicit networks, as well as a combined image fusion and registration objective, to enhance the performance of implicit registration networks beyond the standard approach. The investigated implicit methods are compared to the VoxelMorph convolutional neural network and to the symmetric image normalization (SyN) registration algorithm from the Advanced Normalization Tools (ANTs). Our findings not only highlight the remarkable capabilities of implicit networks in addressing pairwise image registration challenges, but also showcase their potential as a powerful and versatile off-the-shelf tool in the fields of neuroscience and radiology.

We introduce "PatchMorph," an new stochastic deep learning algorithm tailored for unsupervised 3D brain image registration. Unlike other methods, our method uses compact patches of a constant small size to derive solutions that can combine global transformations with local deformations. This approach minimizes the memory footprint of the GPU during training, but also enables us to operate on numerous amounts of randomly overlapping small patches during inference to mitigate image and patch boundary problems. PatchMorph adeptly handles world coordinate transformations between two input images, accommodating variances in attributes such as spacing, array sizes, and orientations. The spatial resolution of patches transitions from coarse to fine, addressing both global and local attributes essential for aligning the images. Each patch offers a unique perspective, together converging towards a comprehensive solution. Experiments on human T1 MRI brain images and marmoset brain images from serial 2-photon tomography affirm PatchMorph's superior performance.

In this work, we investigate the usefulness of vision-language models (VLMs) and large language models for binary few-shot classification of medical images. We utilize the GPT-4 model to generate text descriptors that encapsulate the shape and texture characteristics of objects in medical images. Subsequently, these GPT-4 generated descriptors, alongside VLMs pre-trained on natural images, are employed to classify chest X-rays and breast ultrasound images. Our results indicate that few-shot classification of medical images using VLMs and GPT-4 generated descriptors is a viable approach. However, accurate classification requires to exclude certain descriptors from the calculations of the classification scores. Moreover, we assess the ability of VLMs to evaluate shape features in breast mass ultrasound images. We further investigate the degree of variability among the sets of text descriptors produced by GPT-4. Our work provides several important insights about the application of VLMs for medical image analysis.

We present the first automated pipeline to create an atlas of in situ hybridization gene expression in the adult marmoset brain in the same stereotaxic space. The pipeline consists of segmentation of gene expression from microscopy images and registration of images to a standard space. Automation of this pipeline is necessary to analyze the large volume of data in the genome-wide whole-brain dataset, and to process images that have varying intensity profiles and expression patterns with minimal human bias. To reduce the number of labelled images required for training, we develop a semi-supervised segmentation model. We further develop an iterative algorithm to register images to a standard space, enabling comparative analysis between genes and concurrent visualization with other datasets, thereby facilitating a more holistic understanding of primate brain structure and function.

We investigate the usefulness of formula-driven supervised learning (FDSL) for breast ultrasound (US) image analysis. Medical data are usually too scarce to develop a better performing deep learning model from scratch. Transfer learning with networks pre-trained on ImageNet is commonly applied to address this problem. FDSL techniques have been recently investigated as an alternative solution to ImageNet based approaches. In the FDSL setting, networks for transfer learning applications are developed using large amounts of synthetic images generated with mathematical formulas, possibly taking into account the characteristics of the target data. In this work, we use Field II to develop a large synthetic dataset of 100 000 US images presenting different contour objects, as shape features play an important role in breast mass characterization in US. Synthetic data are utilized to pre-train the ResNet50 classification model and various variants of the U-Net segmentation network. Next, the pre-trained models are fine-tuned on breast mass US images. Our results demonstrate that the proposed FDSL approach can provide good performance with respect to breast mass classification and segmentation.

INTRODUCTION Nonalcoholic fatty liver disease (NAFLD) is a common liver abnormality, but its nonin-vasive diagnosis in patients with severe obesity remains difficult.OBJECTIVES Our aim was to investigate the usefulness of the ultrasound-based hepatorenal index (HRI) technique and 2 biomarker-based methods, including the hepatic steatosis index (HSI) and NAFLD logit score for the diagnosis of NAFLD in patients referred for bariatric surgery.PATIENTS AND METHODS A total of 162 patients, including 106 with NAFLD, admitted for bariatric surgery participated in the study. Fat fraction level and the presence of NAFLD were determined using surgical liver biopsy. Each patient underwent liver ultrasound examination and blood tests to determine the HRI, HSI, and NAFLD logit score.RESULTS For the NAFLD diagnosis, the HRI, HSI, and NAFLD logit score techniques achieved areas under the receiver operating characteristic curves of 0.879, 0.577, and 0.825, respectively. The Spearman cor-relation coefficients between the liver fat fraction values and the HRI, HSI, and NAFLD logit score were equal to 0.695, 0.215, and 0.595, respectively. The optimal cutoff values for the NAFLD diagnosis for the HRI, HSI, and NAFLD logit score were equal to 1.12, 56.1, and 0.59, respectively, and significantly differed from the cutoff values reported for the general population in the literature.CONCLUSIONS Our study confirmed the usefulness of only 2 out of 3 techniques, the HRI and the NAFLD logit score for the diagnosis of NAFLD in patients with severe obesity. The methods designed for the general population require different cutoff values to achieve accurate performance in patients with severe obesity.

We propose a novel image registration method based on implicit neural representations that addresses the challenging problem of registering a pair of brain images with similar anatomical structures, but where one image contains additional features or artifacts that are not present in the other image. To demonstrate its effectiveness, we use 2D microscopy in situ hybridization gene expression images of the marmoset brain. Accurately quantifying gene expression requires image registration to a brain template, which is difficult due to the diversity of patterns causing variations in visible anatomical brain structures. Our approach uses implicit networks in combination with an image exclusion loss to jointly perform the registration and decompose the image into a support and residual image. The support image aligns well with the template, while the residual image captures individual image characteristics that diverge from the template. In experiments, our method provided excellent results and outperformed other registration techniques.

PurposeBUS-Set is a reproducible benchmark for breast ultrasound (BUS) lesion segmentation, comprising of publicly available images with the aim of improving future comparisons between machine learning models within the field of BUS. MethodFour publicly available datasets were compiled creating an overall set of 1154 BUS images, from five different scanner types. Full dataset details have been provided, which include clinical labels and detailed annotations. Furthermore, nine state-of-the-art deep learning architectures were selected to form the initial benchmark segmentation result, tested using five-fold cross-validation and MANOVA/ANOVA with Tukey statistical significance test with a threshold of 0.01. Additional evaluation of these architectures was conducted, exploring possible training bias, and lesion size and type effects. ResultsOf the nine state-of-the-art benchmarked architectures, Mask R-CNN obtained the highest overall results, with the following mean metric scores: Dice score of 0.851, intersection over union of 0.786 and pixel accuracy of 0.975. MANOVA/ANOVA and Tukey test results showed Mask R-CNN to be statistically significant better compared to all other benchmarked models with a p-value >0.01. Moreover, Mask R-CNN achieved the highest mean Dice score of 0.839 on an additional 16 image dataset, that contained multiple lesions per image. Further analysis on regions of interest was conducted, assessing Hamming distance, depth-to-width ratio (DWR), circularity, and elongation, which showed that the Mask R-CNN's segmentations maintained the most morphological features with correlation coefficients of 0.888, 0.532, 0.876 for DWR, circularity, and elongation, respectively. Based on the correlation coefficients, statistical test indicated that Mask R-CNN was only significantly different to Sk-U-Net. ConclusionsBUS-Set is a fully reproducible benchmark for BUS lesion segmentation obtained through the use of public datasets and GitHub. Of the state-of-the-art convolution neural network (CNN)-based architectures, Mask R-CNN achieved the highest performance overall, further analysis indicated that a training bias may have occurred due to the lesion size variation in the dataset. All dataset and architecture details are available at GitHub: , which allows for a fully reproducible benchmark.