Andrew Peter King

Echocardiography is the first imaging modality used for assessing cardiac function, and accurate segmentation of cardiac structures is essential for deriving biomarkers. However, the development of effective automated segmentation models for multiple cardiac structures is challenged by the difficulty of training on datasets from different sources that are often partially-labelled. This study aims to address this challenge by evaluating the performance of three loss functions - adaptive categorical cross entropy (aCCE) loss, marginal loss, and the adaptive binary cross entropy (aBCE) loss - in handling partially-labelled data. We conduct a comprehensive comparison of these loss functions across multiple scenarios and network architectures (R2.1): intra-domain and inter-domain tasks, with both single and multiple partial-labels, and varying proportions of fully-labelled to partially-labelled data.Our experiments reveal that all three loss functions exhibit strong performance in intra- domain segmentation tasks, effectively handling label variations within the same domain. For inter-domain tasks, where models are trained on datasets with a domain shift, the aBCE and marginal losses show superior performance when dealing with the case of one label being missing from some training examples. In scenarios involving more than one label being missing, marginal loss outperforms the other methods, demonstrating its robustness in such complex conditions. These results highlight the strengths of each loss function depending on the labelling scenario, emphasizing the importance of selecting the appropriate loss function to optimize model performance. This study represents the first investigation of techniques for handling partially-labelled data from multiple different domains in echocardiography segmentation and provides a comprehensive comparison of loss-based (R3.4) solutions.

Recent work has shown improved lesion detectability and flexibility to reconstruction hyperparameters (e.g. scanner geometry or dose level) when PET images are reconstructed by leveraging pre-trained diffusion models. Such methods train a diffusion model (without sinogram data) on high-quality, but still noisy, PET images. In this work, we propose a simple method for generating subject-specific PET images from a dataset of multi-subject PET-MR scans, synthesizing “pseudo-PET" images by transforming between different patients’ anatomy using image registration. The images we synthesize retain information from the subject’s MR scan, leading to higher resolution and the retention of anatomical features compared to the original set of PET images. With simulated and real datasets, we show that pre-training a personalized diffusion model with subject-specific “pseudo-PET" images improves reconstruction accuracy with low-count data. In particular, the method shows promise in combining information from a guidance MR scan without overly imposing anatomical features, demonstrating an improved trade-off between reconstructing PET-unique image features versus features present in both PET and MR. We believe this approach for generating and utilizing synthetic data has further applications to medical imaging tasks, particularly because patient-specific PET images can be generated without resorting to generative deep learning or large training datasets.

Deep neural networks often exploit shortcuts. These are spurious cues which are associated with output labels in the training data but are unrelated to task semantics. When the shortcut features are associated with sensitive attributes, shortcut learning can lead to biased model performance. Existing methods for localising and understanding shortcut learning are mostly based upon qualitative, image-level inspection and assume cues are human-visible, limiting their use in domains such as medical imaging. We introduce OSCAR (Ordinal Scoring Correlations for Attribution Representations), a model-agnostic framework for quantifying shortcut learning and localising shortcut features. OSCAR converts image-level task attribution maps into dataset-level rank profiles of image regions and compares them across three models: a balanced baseline model (BA), a test model (TS), and a sensitive attribute predictor (SA). By computing pairwise, partial, and deviation-based correlations on these rank profiles, we produce a set of quantitative metrics that characterise the degree of shortcut reliance for TS, together with a ranking of image-level regions that contribute most to it. Experiments on CelebA, CheXpert, and ADNI show that our correlations are (i) stable across seeds and partitions, (ii) sensitive to the level of association between shortcut features and output labels in the training data, and (iii) able to distinguish localised from diffuse shortcut features. As an illustration of the utility of our method, we show how worst-group performance disparities can be reduced using a simple test-time attenuation approach based on the identified shortcut regions. OSCAR provides a lightweight, pixel-space audit that yields statistical decision rules and spatial maps, enabling users to test, localise, and mitigate shortcut reliance. The code is available at https://github.com/acharaakshit/oscar

Background: Sub-Saharan Africa (SSA) bears the highest global burden of under-5 mortality, with congenital heart disease (CHD) as a major contributor. Despite advancements in high-income countries, CHD-related mortality in SSA remains largely unchanged due to limited diagnostic capacity and centralized health care. While pulse oximetry aids early detection, confirmation typically relies on echocardiography, a procedure constrained by a shortage of specialized personnel. Artificial intelligence (AI) offers a promising solution to bridge this diagnostic gap. Objective: This study aims to develop an AI-assisted echocardiography system that enables nonexpert operators, such as nurses, midwives, and medical doctors, to perform basic cardiac ultrasound sweeps on neonates suspected of CHD and extract accurate cardiac images for remote interpretation by a pediatric cardiologist. Methods: The study will use a 2-phase approach to develop a deep learning model for real-time cardiac view detection in neonatal echocardiography, utilizing data from St. Padre Pio Hospital in Cameroon and the Red Cross War Memorial Children’s Hospital in South Africa to ensure demographic diversity. In phase 1, the model will be pretrained on retrospective data from nearly 500 neonates (0-28 days old). Phase 2 will fine-tune the model using prospective data from 1000 neonates, which include background elements absent in the retrospective dataset, enabling adaptation to local clinical environments. The datasets will consist of short and continuous echocardiographic video clips covering 10 standard cardiac views, as defined by the American Society of Echocardiography. The model architecture will leverage convolutional neural networks and convolutional long short-term memory layers, inspired by the interleaved visual memory framework, which integrates fast and slow feature extractors via a shared temporal memory mechanism. Video preprocessing, annotation with predefined cardiac view codes using Labelbox, and training with TensorFlow and PyTorch will be performed. Reinforcement learning will guide the dynamic use of feature extractors during training. Iterative refinement, informed by clinical input, will ensure that the model effectively distinguishes correct from incorrect views in real time, enhancing its usability in resource-limited settings. Results: Retrospective data collection for the project began in September 2024, and to date, data from 308 babies have been collected and labeled. In parallel, the initial model framework has been developed and training initiated using a subset of the labeled data. The project is currently in the intensive execution phase, with all objectives progressing in parallel and final results expected within 10 months. Conclusions: The AI-assisted echocardiography model developed in this project holds promise for improving early CHD diagnosis and care in SSA and other low-resource settings. International Registered Report Identifier (IRRID): DERR1-10.2196/75270

Background and Aims: Cardiovascular ageing is a progressive loss of physiological reserve, modified by environmental and genetic risk factors, that contributes to multi-morbidity due to accumulated damage across diverse cell types, tissues, and organs. Obesity is implicated in premature ageing, but the effect of body fat distribution in humans is unknown. This study determined the influence of sex-dependent fat phenotypes on human cardiovascular ageing. Methods: Data from 21 241 participants in the UK Biobank were analysed. Machine learning was used to predict cardiovascular age from 126 image-derived traits of vascular function, cardiac motion, and myocardial fibrosis. An age-delta was calculated as the difference between predicted age and chronological age. The volume and distribution of body fat was assessed from whole-body imaging. The association between fat phenotypes and cardiovascular age-delta was assessed using multivariable linear regression with age and sex as co-covariates, reporting β coefficients with 95% confidence intervals (CI). Two-sample Mendelian randomization was used to assess causal associations. Results: Visceral adipose tissue volume [β = 0.656, (95% CI, .537–.775), P < .0001], muscle adipose tissue infiltration [β = 0.183, (95% CI, .122–.244), P = .0003], and liver fat fraction [β = 1.066, (95% CI .835–1.298), P < .0001] were the strongest predictors of increased cardiovascular age-delta for both sexes. Abdominal subcutaneous adipose tissue volume [β = 0.432, (95% CI, .269–.596), P < .0001] and android fat mass [β = 0.983, (95% CI, .64–1.326), P < .0001] were each associated with increased age-delta only in males. Genetically predicted gynoid fat showed an association with decreased age-delta. Conclusions: Shared and sex-specific patterns of body fat are associated with both protective and harmful changes in cardiovascular ageing, highlighting adipose tissue distribution and function as a key target for interventions to extend healthy lifespan.

Artificial intelligence (AI) is increasingly being used for medical imaging tasks. However, there can be biases in AI models, particularly when they are trained using imbalanced training datasets. One such example has been the strong ethnicity bias effect in cardiac magnetic resonance (CMR) image segmentation models. Although this phenomenon has been reported in a number of publications, little is known about the effectiveness of bias mitigation algorithms in this domain. We aim to investigate the impact of common bias mitigation methods to address bias between Black and White subjects in AI-based CMR segmentation models. Specifically, we use oversampling, importance reweighing and Group DRO as well as combinations of these techniques to mitigate the ethnicity bias. Second, motivated by recent findings on the root causes of AI-based CMR segmentation bias, we evaluate the same methods using models trained and evaluated on cropped CMR images. We find that bias can be mitigated using oversampling, significantly improving performance for the underrepresented Black subjects whilst not significantly reducing the majority White subjects’ performance. Using cropped images increases performance for both ethnicities and reduces the bias, whilst adding oversampling as a bias mitigation technique with cropped images reduces the bias further. When testing the models on an external clinical validation set, we find high segmentation performance and no statistically significant bias.

Background Echocardiography after myocardial infarction (MI) provides clinically-useful information through assessment of regional wall motion abnormalities, but interpretation requires expertise and remains subject to intra- and inter-observer variability. Artificial intelligence (AI) has shown promise in the automatic interpretation of echocardiography images, but it is unclear how explainability of the AI model output will change the overall collaborative performance of the human-AI team. Methods A ResNet18-LSTM model with attention was trained to classify normal vs MI on 127 apical four chamber (A4C) and 120 apical two chamber (A2C) echocardiogram videos from the HMC-QU dataset. Gradient-weighted Class Activation Mapping (GradCAM) was used for visual explanations. An end-user study with eight cardiology trainees compared diagnostic performance across three conditions: (a) echo clips alone, (b) echo clips with AI predictions, and (c) echo clips with AI predictions plus GradCAM. explanations. Results The AI models demonstrated strong discriminative performance with AUCs of 0.9429 (A2C) and 0.9250 (A4C). The end user study revealed that AI alone achieved 80.0% accuracy versus 77.0% for clinicians alone. Surprisingly, combining AI with human judgement did not improve performance, and introducing visual explanations reduced accuracy to 72% and specificity from 93.8% to 83.8% (p = 0.046). Conclusion Our findings suggest that while AI models can effectively detect MI on echocardiographic videos, current explainability techniques may misalign with clinical reasoning, potentially impairing diagnostic performance. Future effective integration of AI into echocardiographic workflows will require not only high accuracy but also AI visual explanation strategies that complement clinician expertise.