Julia Alison Noble

Background: In this study, we use quantitative and qualitative assessments to evaluate the feasibility and usability of an obstetric ultrasound video colour overlay to assist placenta localisation and hence risk stratify high-risk pregnancies that have a low-lying placenta. The colour overlay (designed in a previous study) simplifies clinical interpretation of ultrasound video to support screening for pregnant women in rural and low- and middle-income country (LMIC) settings who do not receive care, partly because of a trained operator shortage. Methods: For feasibility, we assessed risk stratification (or checking for a low-lying placenta) of a trained sonographer (6.5 years of experience) using the combined video and colour overlay compared with a hospital freehand scan without a colour overlay. For usability, we recruited 31 Indian healthcare system clinicians in both north and south India hospital sites. We assessed clinical opinions and perceptions of the colour overlay over four study rounds of written questionnaires and a single round of oral interviews. In each round, participants interacted with 20 colour overlays on a portable tablet computer. Results: For feasibility, the sonographer successfully risk stratified the 2 low-lying, high-risk placentas out of a cohort of 55 unique pregnancies (1 video per pregnancy). However, this approach also resulted in 18 false positives (normal placentas that were incorrectly labelled as low-lying). Future feasibility assessment would need to use a larger cohort of pregnancies to strengthen the preliminary findings reported here. For usability, participants gave positive written responses and reported an increase in both confidence when using the colour overlay, and trust in the colour overlay post-study compared to pre-study. From qualitative coding of interview transcripts (1,643 codes and 13 themes generated), the colour overlay was overwhelmingly perceived as an effective risk stratification tool and usable by healthcare workers in rural and LMIC settings. Non-trivial perceived challenges to translation included compliance of the colour overlay with India’s current fetal sex determination law (Pre-Conception and Pre-Natal Diagnostic Techniques Act, 1994), including compliance of targeted users of rural healthcare workers (who cannot currently legally use ultrasound devices), and ultrasound device purchase and maintenance costs. Conclusions: Overall, the colour overlay was perceived favourably by clinicians. For successful adoption of the technology in the Indian healthcare system, the non-technical issues raised would need to be carefully and sensitively addressed.

Objective Structured Clinical Examinations (OSCEs) are typically scored by a single examiner per station, exposing results to hard-to-audit examiner variability. We developed the Rater-Aware Verification Network (RAVEN), a multimodal Artificial Intelligence (AI) system fusing egocentric video, examiner verbalisations and marks, and virtual reality (VR) action logs as an examiner-conditioned second marker for paediatric OSCEs (retrospective evaluation; 120 students, 442 ratings, eight domains). A confidence-gated hybrid improved agreement with leave-one-examiner-out consensus on the pass/fail decision (AC1: +6.2 percentage points; p = 0.0004) and domain scores (mean AC2: +3.0 points, five of eight significant), with the largest gains on borderline-fail cases (+16.3 points of concordance with consensus). Comparing AI-inferred, rubric-specified, and examiner-articulated criteria revealed implicit practices absent from marking guidelines. Agreement is measured against a panel-derived reference rather than an external ground truth; the system is intended as an audit and flagging aid for human adjudication rather than an autonomous decision-maker.

Schistosomiasis causes a complex, difficult to diagnose form of liver fibrosis with high rates of life-threatening morbidity in resource-poor settings where there are often no trained sonographers. Protocols for diagnosis of schistosomiasis-related liver fibrosis have focused on difficult-to-acquire and subjective ultrasound images dependent on extensive expertise. Here we present SchistoTrackVideoNet, the first deep learning-based video model trained on easy-to-acquire standardised ultrasound video sweeps for classification of schistosomiasis-related liver fibrosis. This video-based classification model was trained and evaluated on video sweeps from 2140 participants aged 5--87 years from three districts in rural Uganda. We tested the model at a clinically-relevant sensitivity threshold ($\geq$90\%) and achieved positive predictive values of 0.0968--0.5556 for diverse presentations of liver fibrosis. Our findings show potential for the use of easy-to-acquire video sweeps for diagnosis of schistosomiasis-related liver fibrosis and our model provides a proof-of-concept for deep learning applied to liver ultrasound video for diagnosis of schistosomiasis-related liver morbidity.

Liver fibrosis is a major cause of death in low- and middle-income country contexts. In rural, poor areas of sub-Saharan Africa, schistosomiasis is an underestimated cause of liver fibrosis. Despite the need for increased diagnostic capacity for schistosomiasis-related liver fibrosis, there are no automated, clinically-validated tools to diagnose schistosomiasis-related liver fibrosis. We present SchistoTrackNet which is, to our knowledge, the first deep learning-based model for distinguishing distinct presentations of schistosomiasis-related liver fibrosis of varying severity. Ultrasound images from 1533 participants aged 5–84 years from three districts in rural Uganda were used to train and evaluate the presented models. The models were evaluated by assessing failure cases and by comparing results with re-readings performed by sonographers experienced in diagnosis of schistosomiasis morbidity. Our models show potential to enable automated reading of ultrasound images for schistosomiasis-related liver fibrosis to allow large-scale surveillance of schistosomiasis morbidity and contribute towards the World Health Organization target to eliminate schistosomiasis as a public health problem.

We previously constructed a qualitative, 3D ultrasound derived atlas of the normative spatiotemporal dynamics of fetal brain maturation. Here, using the same healthy multi-national cohort, we applied deep learning methods to 4205 fetal brain scans from 18–27 weeks’ gestation, to produce an extensive, quantitative description of the growth of 16 fetal brain structures associated with satisfactory domain-specific neurodevelopmental scores at 2 years of age. The methodology, which is publicly available, takes less than 10 seconds per scan. We define 28 region-specific, functionally relevant, normative growth trajectories, a ratio between the relative volumes of the insular (rILV) and parietal (rPLV) lobes reflecting asynchronous maturation of fetal brain regions, and introduce a fetal brain maturation index that quantifies biological age and deviations from chronological age. Finally, the very low percentage of variance explained by between site differences (0.6% to 5.8% of the total variance) reinforces a fundamental biological principle: fetal growth and development across populations with diverse ancestries is similar provided that environmental constraints on growth are minimal.

Freehand fetal ultrasound examinations require substantial clinical skill. Here we propose Sonomate (mate of a sonographer), an AI assistant to a user during fetal ultrasound examinations. Sonomate is based on aligning video features and text features derived from transcribed audio to facilitate real-time interactions between an ultrasound machine and a user. Our approach combines coarse-grained video–text alignment with fine-grained image–sentence alignment to build a robust visually grounded language model capable of understanding fetal ultrasound videos. To tackle the challenges associated with heterogeneous language and asynchronous content in real-world video–audio pairs, we design the anatomy-aware alignment and context label correction in the fine-grained alignment. Sonomate is effective at anatomy detection in fetal ultrasound images without the need for retraining on manually annotated data. Furthermore, Sonomate shows promising performance in visual question answering for both fetal ultrasound images and videos. Guardrails are built to ensure the safety of Sonomate during deployment. This advancement paves the way towards AI-assistive technology being used to support sonography training and enhanced diagnostic capabilities.

Explainable Artificial Intelligence (XAI) is proposed as essential for high-risk applications like healthcare, where it aims to enhance user trust. However, studies often rely on automated metrics rather than user evaluation. We adapt a prototype-based XAI model for image-based gestational age (GA) estimation and evaluate its impact on trust, reliance, and performance, including a novel measure of appropriate reliance. Ten sonographers completed a 3-stage reader study assessing the XAI model’s impact on GA estimates. Model predictions reduced clinician mean absolute error (MAE) from 23.5 to 15.7 days, and explanations had a further non-significant reduction to 14.3 days. However, the impact of explanations varied across participants, with some performing worse with explanations than without. Additionally, although explanations increased participant confidence, they had no significant effect on trust or reliance on the model. These counterintuitive results highlight potential pitfalls in deploying XAI, emphasising the need for human studies to capture clinician variability.