Ultrasound (US) imaging is clinically invaluable due to its non-invasive and safe nature. However, interpreting US images is challenging, requires significant expertise, and time, and is often prone to errors. Deep learning offers assistive solutions such as segmentation. Supervised methods rely on large, high-quality, and consistently labelled datasets, which are challenging to curate. Moreover, these methods tend to underperform on out-of-distribution data, limiting their clinical utility. Self-supervised learning (SSL) has emerged as a promising alternative, leveraging unlabelled data to enhance model performance and generalisability. We introduce a contrastive SSL approach tailored for B-mode US images, incorporating a novel Relation Contrastive Loss (RCL). RCL encourages learning of distinct features by differentiating positive and negative sample pairs through a learnable metric. Additionally, we propose spatial and frequency-based augmentation strategies for the representation learning on US images.Our approach significantly outperforms traditional supervised segmentation methods across three public breast US datasets, particularly in data-limited scenarios. Notable improvements on the Dice similarity metric include a 4% increase on 20% and 50% of the BUSI dataset, nearly 6% and 9% improvements on 20% and 50% of the BrEaST dataset, and 6.4% and 3.7% improvements on 20% and 50% of the UDIAT dataset, respectively. Furthermore, we demonstrate superior generalisability on the out-of-distribution UDIAT dataset with performance boosts of 20.6% and 13.6% compared to the supervised baseline using 20% and 50% of the BUSI and BrEaST training data, respectively. Our research highlights that domain-inspired SSL can improve US segmentation, especially under data-limited conditions.

BACKGROUND:Glioblastoma (GBM) presents a significant clinical challenge due to its aggressive nature and extensive heterogeneity. Tumor purity, the proportion of malignant cells within a tumor, is an important covariate for understanding the disease, having direct clinical relevance or obscuring signal of the malignant portion in molecular analyses of bulk samples. However, current methods for estimating tumor purity are nonspecific and technically demanding. Therefore, we aimed to build a reliable and accessible purity estimator for GBM. METHODS:We developed GBMPurity, a deep learning model specifically designed to estimate the purity of IDH-wild type primary GBM from bulk RNA-sequencing (RNA-seq) data. The model was trained using simulated pseudobulk tumors of known purity from labeled single-cell data acquired from the GBmap resource. The performance of GBMPurity was evaluated and compared to several existing tools using independent datasets. RESULTS:GBMPurity outperformed existing tools, achieving a mean absolute error of 0.15 and a concordance correlation coefficient of 0.88 on validation datasets. We demonstrate the utility of GBMPurity through inference on bulk RNA-seq samples and observe reduced purity of the proneural molecular subtype relative to the classical, attributed to the increased presence of healthy brain cells. CONCLUSIONS:GBMPurity provides a reliable and accessible tool for estimating tumor purity from bulk RNA-seq data, enhancing the interpretation of bulk RNA-seq data and offering valuable insights into GBM biology. To facilitate the use of this model by the wider research community, GBMPurity is available as a web-based tool at: https://gbmdeconvoluter.leeds.ac.uk/.

Generative modelling for shapes is a prerequisite for In-Silico Clinical Trials (ISCTs), which aim to cost-effectively validate medical device interventions using synthetic anatomical shapes, often represented as 3D surface meshes. However, constructing AI models to generate shapes closely resembling the real mesh samples is challenging due to variable vertex counts, connectivities, and the lack of dense vertex-wise correspondences across the training data. Employing graph representations for meshes, we develop a novel unsupervised geometric deep-learning model to establish refinable shape correspondences in a latent space, construct a population-derived atlas and generate realistic synthetic shapes. We additionally extend our proposed base model to a joint shape generative-clustering multi-atlas framework to incorporate further variability and preserve more details in the generated shapes. Experimental results using liver and left-ventricular models demonstrate the approach's applicability to computational medicine, highlighting its suitability for ISCTs through a comparative analysis.

Colonoscopy screening is the gold standard procedure for assessing abnormalities in the colon and rectum, such as ulcers and cancerous polyps. Measuring the abnormal mucosal area and its 3D reconstruction can help quantify the surveyed area and objectively evaluate disease burden. However, due to the complex topology of these organs and variable physical conditions, for example, lighting, large homogeneous texture, and image modality estimating distance from the camera (aka depth) is highly challenging. Moreover, most colonoscopic video acquisition is monocular, making the depth estimation a non-trivial problem. While methods in computer vision for depth estimation have been proposed and advanced on natural scene datasets, the efficacy of these techniques has not been widely quantified on colonoscopy datasets. As the colonic mucosa has several low-texture regions that are not well pronounced, learning representations from an auxiliary task can improve salient feature extraction, allowing estimation of accurate camera depths. In this work, we propose to develop a novel multi-task learning (MTL) approach with a shared encoder and two decoders, namely a surface normal decoder and a depth estimator decoder. Our depth estimator incorporates attention mechanisms to enhance global context awareness. We leverage the surface normal prediction to improve geometric feature extraction. Also, we apply a cross-task consistency loss among the two geometrically related tasks, surface normal and camera depth. We demonstrate an improvement of 15.75% on relative error and 10.7% improvement on δ1.25 accuracy over the most accurate baseline state-of-the-art Big-to-Small (BTS) approach. All experiments are conducted on a recently released C3VD dataset, and thus, we provide a first benchmark of state-of-the-art methods on this dataset.

Ultrasound imaging faces particular challenges with high inter-operator variability and manual inspection of abnormalities. Deep learning segmentation methods are progressing rapidly to address these clinical challenges with improved automatic segmentation performance combining convolutional neural network (CNN) and Transformer approaches. However, challenges still remain with poor performance in boundary areas due to high speckle noise and training models with limited training data. This paper demonstrates a comparison of EfficientNet B2, EfficientNet B7 and PVT-v2-B5 pre-trained encoder backbones in a U-Net architecture. A noticeable improvement for all three pre-trained backbones is shown particularly in smaller Breast Ultrasound datasets with limited training data (EfficientNet B2: 4.5%, EfficientNet B7: 3.9%, PVT-v2-B5: 2.4%). However, the improvement is marginal (less than 1%) in the larger Nerve Ultrasound dataset. In addition, we noticed that the performance across all backbones is better in segmenting regular regions of interest (e.g. benign breast lesions), over irregular shapes (e.g. malignant breast lesions). The code used for this study is available at: https://github.com/aimsgroup-Leeds/IEEECBMS2024_US_Seg

Purpose To review methodological approaches for automated segmentation of subcutaneous adipose tissue, visceral adipose tissue, and skeletal muscle from abdominal cross-sectional imaging for body composition analysis. Method Four databases were searched for publications describing automated segmentation of subcutaneous adipose tissue, visceral adipose tissue, and/or skeletal muscle from abdominal CT or MR imaging between 2019 and 2023. Included reports were evaluated to assess how imaging modality, cohort size, vertebral level, model dimensionality, and use of a volume or single slice affected segmentation accuracy and/or clinical utility. Exclusion criteria included reports not in English language, manual or semi-automated segmentation methods, reports prior to 2019 or solely of paediatric patients, and those not describing the use of abdominal CT or MR. Results After exclusions, 172 reports were included in the review. CT imaging was utilised approximately four times as often as MRI, and segmentation accuracy did not significantly differ between the two modalities. Cohort size had no significant effect on segmentation accuracy. There was little evidence to refute the current practice of extracting body composition metrics from the third lumbar vertebral level. There was no clear benefit of using a 3D model to perform segmentation over a 2D approach. Conclusion Automated segmentation of intra-abdominal soft tissues for body composition analysis is an intense area of research activity. Segmentation accuracy is not affected by cross-sectional imaging modality. Extracting metrics from a single slice at the third lumbar vertebral level is a common approach, however, extracting metrics from a volumetric slab surrounding this level may increase the resilience of the technique, which is important for clinical translation. A paucity of publicly available datasets led to most reports using different data sources, preventing direct comparison of segmentation techniques. Future efforts should prioritise creating a standardised dataset to facilitate benchmarking of different algorithms and subsequent clinical adoption.

While several previous studies have devised methods for segmentation of polyps, most of these methods are not rigorously assessed on multi-center datasets. Variability due to appearance of polyps from one center to another, difference in endoscopic instrument grades, and acquisition quality result in methods with good performance on in-distribution test data, and poor performance on out-of-distribution or underrepresented samples. Unfair models have serious implications and pose a critical challenge to clinical applications. We adapt an implicit bias mitigation method which leverages Bayesian epistemic uncertainties during training to encourage the model to focus on underrepresented sample regions. We demonstrate the potential of this approach to improve generalisability without sacrificing state-of-the-art performance on a challenging multi-center polyp segmentation dataset (PolypGen) with different centers and image modalities.

Abstract AIMS The rapid recurrence of therapy-resistant glioblastoma leaves patients with a dismal 2-year survival rate of just 16%, highlighting the need to develop treatments that combat or prevent glioblastoma recurrence. Recently, glioblastoma patients have been stratified by their gene expression response to therapy. Specifically, genes with JARID2 binding sites in their promoter (JBS-genes) either become up- or down-regulated on progression from the primary to the recurrent tumour. Importantly, these two responses appear to employ different mechanisms of therapy resistance. Using gene expression data from primary tumours, we aim to predict the direction of JBS- gene dysregulation. Ultimately, this prediction could guide treatment decisions that prevent the acquisition of therapy resistance in glioblastoma. METHOD We used gene expression data from 84 primary glioblastoma tumours that had been labelled as either ‘Up’ or ‘Down’ responders based on the direction of JBS-gene dysregulation in matched recurrent tumours. We converted the RNA-sequencing data from the primary tumours to single-sample pathway scores, which were used to train several binary classification machine learning models. RESULTS A ridge regression model achieved an accuracy of 73.9% in predicting the direction of JBS-gene dysregulation in an independent set of 23 primary glioblastoma tumours. CONCLUSIONS Our classifier performs better than random chance demonstrating that gene expression data is informative and molecular features influence this gene expression response to therapy. However, there remains scope for improvement. This suggests further studies may benefit from integrating multiple data types such as genomics, epigenomics, tumour architecture, or clinical variables.

Generative statistical models have a wide range of applications in the modelling of anatomies. In-silico clinical trials of medical devices, for instance, require the development of virtual populations of anatomy that capture enough variability while remaining plausible. Model construction and use are heavily influenced by the correspondence problem and establishing shape matching over a large number of training data.This study focuses on generating virtual cohorts of left ventricle geometries resembling different-sized shape populations, suitable for in-silico experiments. We present an unsupervised data-driven probabilistic generative model for shapes. This framework incorporates an attention-based shape matching procedure using graph neural networks, coupled with a $\beta-$VAE generation model, eliminating the need for initial shape correspondence. Left ventricle shapes derived from cardiac magnetic resonance images available in the UK Biobank are utilized for training and validating the framework. We investigate our method’s generative capabilities in terms of generalisation and specificity and show that it is able to synthesise virtual populations of realistic shapes with volumetric measurements in line with actual clinical indices. Moreover, results show our method outperforms joint registration-PCA-based models.

The fairness of a deep neural network is strongly affected by dataset bias and spurious correlations, both of which are usually present in modern feature-rich and complex visual datasets. Due to the difficulty and variability of the task, no single de-biasing method has been universally successful. In particular, implicit methods not requiring explicit knowledge of bias variables are especially relevant for real-world applications. We propose a novel implicit mitigation method using a Bayesian neural network, allowing us to leverage the relationship between epistemic uncertainties and the presence of bias or spurious correlations in a sample. Our proposed posterior estimate sharpening procedure encourages the network to focus on core features that do not contribute to high uncertainties. Experimental results on three benchmark datasets demonstrate that Bayesian networks with sharpened posterior estimates perform comparably to prior existing methods and show potential worthy of further exploration.

We propose a novel method for generating high-resolution videos of talking-heads from speech audio and a single 'identity' image. Our method is based on a convolutional neural network model that incorporates a pre-trained StyleGAN generator. We model each frame as a point in the latent space of StyleGAN so that a video corresponds to a trajectory through the latent space. Training the network is in two stages. The first stage is to model trajectories in the latent space conditioned on speech utterances. To do this, we use an existing encoder to invert the generator, mapping from each video frame into the latent space. We train a recurrent neural network to map from speech utterances to displacements in the latent space of the image generator. These displacements are relative to the back-projection into the latent space of an identity image chosen from the individuals depicted in the training dataset. In the second stage, we improve the visual quality of the generated videos by tuning the image generator on a single image or a short video of any chosen identity. We evaluate our model on standard measures (PSNR, SSIM, FID and LMD) and show that it significantly outperforms recent state-of-the-art methods on one of two commonly used datasets and gives comparable performance on the other. Finally, we report on ablation experiments that validate the components of the model. The code and videos from experiments can be found at https://mohammedalghamdi.github.io/talking-heads-acm-mm

Differentiable physics modeling combines physics models with gradient-based learning to provide model explicability and data efficiency. It has been used to learn dynamics, solve inverse problems and facilitate design, and is at its inception of impact. Current successes have concentrated on general physics models such as rigid bodies, deformable sheets, etc., assuming relatively simple structures and forces. Their granularity is intrinsically coarse and therefore incapable of modelling complex physical phenomena. Fine-grained models are still to be developed to incorporate sophisticated material structures and force interactions with gradient-based learning. Following this motivation, we propose a new differentiable fabrics model for composite materials such as cloths, where we dive into the granularity of yarns and model individual yarn physics and yarn-to-yarn interactions. To this end, we propose several differentiable forces, whose counterparts in empirical physics are indifferentiable, to facilitate gradient-based learning. These forces, albeit applied to cloths, are ubiquitous in various physical systems. Through comprehensive evaluation and comparison, we demonstrate our model's explicability in learning meaningful physical parameters, versatility in incorporating complex physical structures and heterogeneous materials, data-efficiency in learning, and high-fidelity in capturing subtle dynamics.

Deep neural networks are highly susceptible to learning biases in visual data.While various methods have been proposed to mitigate such bias, the majority require explicit knowledge of the biases present in the training data in order to mitigate.We argue the relevance of exploring methods which are completely ignorant of the presence of any bias, but are capable of identifying and mitigating them.Furthermore, we propose using Bayesian neural networks with an epistemic uncertainty-weighted loss function to dynamically identify potential bias in individual training samples and to weight them during training.We find a positive correlation between samples subject to bias and higher epistemic uncertainties.Finally, we show the method has potential to mitigate visual bias on a bias benchmark dataset and on a real-world face detection problem, and we consider the merits and weaknesses of our approach.

In this paper, we present the first extended catalogue of far-infrared fluxes of Galactic bubbles. Fluxes were estimated for 1814 bubbles, defined here as the ‘golden sample’, and were selected from the Milky Way Project First Data Release (Simpson et al.) The golden sample was comprised of bubbles identified within the Wide-field Infrared Survey Explorer (WISE) dataset (using 12- and 22-μm images) and Herschel data (using 70-, 160-, 250-, 350- and 500-μm wavelength images). Flux estimation was achieved initially via classical aperture photometry and then by an alternative image analysis algorithm that used active contours. The accuracy of the two methods was tested by comparing the estimated fluxes for a sample of bubbles, made up of 126 H ii regions and 43 planetary nebulae, which were identified by Anderson et al. The results of this paper demonstrate that a good agreement between the two was found. This is by far the largest and most homogeneous catalogue of infrared fluxes measured for Galactic bubbles and it is a step towards the fully automated analysis of astronomical datasets.

Process mining techniques can play a significant role in understanding healthcare processes by supporting analysis of patient records in electronic health record systems. Healthcare processes are complex and patterns of care may vary considerably within similar cohorts of patients. Process mining often creates "spaghetti" models and require significant domain expert input to refine. Machine learning approaches such as Hidden Markov Models (HMM) may assist this refinement process. HMMs have been advocated for patient pathways clustering purposes; however these models can also be utilized for detecting hidden processes to help event abstraction. We explore use of an unsupervised method for detecting hidden healthcare sub-processes using HMMs, in particular the Viterbi algorithm. We describe an approach to enrich the event log with HMM-derived states and remodeling the healthcare processes as state transitions using a process mining tool. Our method is applied to event data for 'Altered Mental Status' patients that was extracted from a US hospital database (MIMIC-III). The results are promising and show a successful reduction of model complexity and detection of several hidden processes unsupervised by a domain expert.

Clinical pathways are highly variable and although many patients may follow similar pathway each individual will experience a unique set of events, for example with multiple repeated activities or varied sequences of activities. Process mining techniques are able to discover generalizable pathways based on data mining of event logs but using process mining techniques on a raw clinical pathway data to discover underlying healthcare processes is challenging due to this high variability. This paper involves two main contributions to healthcare process mining. The first contribution is developing a novel approach for event selection and outlier removing in order to improve pattern detection and thus representational quality. The second contribution is to demonstrate a new open access medical dataset, the MIMIC-III (Medical Information Mart for Intensive Care) database, which has not been used in process mining publications. In this paper, we developed a new method for variations reduction in clinical pathways data. Variation can result from outlier events that prevent capturing clear patterns. Our approach targets the behavior of repeated activities. It uses interval-based patterns to determine outlier threshold based on the time of events occurring and the distinctive attribute of observed events. The approach is tested on clinical pathways data for diabetes patients with congestive heart failure extracted from the MIMIC-III medical database and analyzed using the ProM process mining tool. The method has improved model precision conformance without reducing model fitness. We were able to reduce the number of events while making sure the mainstream patterns were unaffected. We found that some activity types had a large number of outlier events whereas other activities had a relatively few. The interval-based event selection method has the potential of improve process visualization. This approach is undergoing implementation as an event log enhancement technique in the ProM tool.

We present a comprehensive statistical analysis of star-forming objects located in the vicinities of 1360 bubble structures throughout the Galactic plane and their local environments. The compilation of ~70 000 star-forming sources, found in the proximity of the ionized (Hii) regions and detected in both Hi-GAL and GLIMPSE surveys, provided a broad overview of the different evolutionary stages of star-formation in bubbles, from prestellar objects to more evolved young stellar objects (YSOs). Surface density maps of star-forming objects clearly reveal an evolutionary trend where more evolved star-forming objects (Class II YSO candidates) are found spatially located near the center, while younger star-forming objects are found at the edge of the bubbles. We derived dynamic ages for a subsample of 182 H ii regions for which kinematic distances and radio continuum flux measurements were available. We detect approximately 80% more star-forming sources per unit area in the direction of bubbles than in the surrounding fields. We estimate the clump formation efficiency (CFE) of Hi-GAL clumps in the direction of the shell of the bubbles to be ~15%, around twice the value of the CFE in fields that are not affected by feedback effects. We find that the higher values of CFE are mostly due to the higher CFE of protostellar clumps, in particular in younger bubbles, whose density of the bubble shells is higher. We argue that the formation rate from prestellar to protostellar phase is probably higher during the early stages of the (H ii ) bubble expansion. Furthermore, we also find a higher fraction of massive YSOs (MYSOs) in bubbles at the early stages of expansion (<2 Myr) than older bubbles. Evaluation of the fragmentation time inside the shell of bubbles advocates the preexistence of clumps in the medium before the bubble expansion in order to explain the formation of MYSOs in the youngest H ii regions (<1 Myr), as supported by numerical simulations. Approximately 23% of the Hi-GAL clumps are found located in the direction of a bubble, with 15% for prestellar clumps and 41% for protostellar clumps. We argue that the high fraction of protostellar clumps may be due to the acceleration of the star-formation process cause by the feedback of the (Hii) bubbles.

To date, the outcomes of cartilage repair have been inconsistent and have frequently yielded mechanically inferior fibrocartilage, thereby increasing the chances of damage recurrence. Implantation of constructs with biochemical composition and mechanical properties comparable to natural cartilage could be advantageous for long-term repair. This study attempted to create such constructs, in vitro, using tissue engineering principles. Bovine synoviocytes were seeded on nonwoven polyethylene terephthalate fiber scaffolds and cultured in chondrogenic medium for 4 weeks, after which uniaxial compressive loading was applied using an in-house bioreactor for 1 h per day, at a frequency of 1 Hz, for a further 84 days. The initial loading conditions, determined from the mechanical properties of the immature constructs after 4 weeks in chondrogenic culture, were strains ranging between 13% and 23%. After 56 days (sustained at 84 days) of loading, the constructs were stained homogenously with Alcian blue and for type-II collagen. Dynamic compressive moduli were comparable to the high end values for native cartilage and proportional to Alcian blue staining intensity. We suggest that these high moduli values were attributable to the bioreactor setup, which caused the loading regime to change as the constructs developed, that is, the applied stress and strain increased with construct thickness and stiffness, providing continued sufficient cell stimulation as further matrix was deposited. Constructs containing cartilage-like matrix with response to load similar to that of native cartilage could produce long-term effective cartilage repair when implanted.

Light microscopy applied to the domain of histopathology has traditionally been a two-dimensional imaging modality. Several authors, including the authors of this work, have extended the use of digital microscopy to three dimensions by stacking digital images of serial sections using image-based registration. In this paper, we give an overview of our approach, and of extensions to the approach to register multi-modal data sets such as sets of interleaved histopathology sections with different stains, and sets of histopathology images to radiology volumes with very different appearance. Our approach involves transforming dissimilar images into a multi-channel representation derived from co-occurrence statistics between roughly aligned images.

Study Design. An in vitro study of the intervertebral disc (IVD) structure using 9.4T magnetic resonance imaging (MRI). Objective. Investigate the potential of ultrahigh-field strength MRI for higher quality 3-dimensional (3D) volumetric MRI datasets of the IVD to better distinguish structural details. Summary of Background Data. MRI has the advantages of being nondestructive and 3D in comparison to most techniques used to obtain the structural details of biological tissues, however, its poor image quality at higher resolution is a limiting factor. Ultrahigh-field MRI could improve the imaging of biological tissues but the current understanding of its application for spinal tissue is limited. Methods. 2 ovine spinal segments (C7–T1, T2–T3) containing the IVD were separately imaged using 2 sequences; 3D spin echo (multislice-multiecho) pulse sequence for the C7–T1 sample and 3D gradient echo (fast-low-angle-shot) pulse sequence for the T2–T3 sample. The C7–T1 sample was subsequently decalcified and imaged again using the same scanning parameters. Histological sections obtained from the decalcified sample were stained followed by digital scanning. Observations from corresponding MRI slices and histological sections were compared as a method of confirmation of morphology captured under MRI. The signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR) and relative-contrast values were calculated for quantitative evaluation of image quality. Results. Measurements from histology sections and corresponding MRI slices matched well. Both sequences revealed finer details of the IVD structure. Under the spin echo sequence, the annulus lamellae architecture was distinguishable and the SNR and CNR values were higher. The relative contrast was considerably higher between high (nucleus) and low (bone) signal constituents, but between the nucleus and the annulus the relative contrast was low. Under the gradient echo sequence, although the relative contrasts between constituents were poor, the fiber orientation was clearly manifested. Conclusion. The obtained positive results demonstrate the potential of ultrahigh-field strength MRI to nondestructively capture the IVD structure. Level of Evidence: N/A