A multi-task deep learning framework assesses echocardiographic image quality and gives real-time guidance on probe positioning. Described in the IEEE Journal of Biomedical and Health Informatics, it combines cardiac view classification, structure identification and corrective movement recommendations. The system is intended to help novice and trainee users acquire acceptable images more consistently.
Shared Network Provides Structure-Specific Feedback
The framework brings several imaging tasks into one shared network. It first determines which standard echocardiographic view is being acquired and segments the visible cardiac structures. The resulting probability outputs are then used to calculate uncertainty. Greater uncertainty indicates that the image is further from an acceptable standard, while lower uncertainty corresponds to clearer and more complete visualisation.
Must Read: LLMs Raise Opportunities and Safeguards for Radiology
The quality measure combines information from view classification, anatomical segmentation, individual structures and their position within the image. This allows the system to move beyond a single overall score. It can indicate whether a chamber, valve or wall is contributing to reduced image quality, giving the operator more specific feedback about what needs correction.
Probe guidance uses features already produced by the shared network rather than analysing the ultrasound image through a separate large model. The system selects among seven actions, including holding the probe or adjusting its tilt, rocking direction or rotation. One model covers eight commonly used views, reducing the need for separate networks for different imaging planes.
The design also avoids an additional round of expert quality scoring. Instead of learning from subjective labels assigned to every image, it derives the assessment from tasks that already use expert-labelled view and structure data. This limits the annotation burden while preserving feedback linked to visible anatomy. The shared design also reduces duplicated processing and supports use on portable devices.
Quality Measure Detects Degraded Cardiac Views
The guidance system was developed and evaluated using a prospectively collected dataset from South Korea. Echocardiographic examinations were performed in adults without a history of heart disease, with an experienced cardiologist deliberately moving the probe from standard positions towards nonstandard views. After exclusions, the analysis included material from 41 participants and covered eight standard views.
The quality-assessment component used these images only for validation because it did not require separate training on quality labels. Across all included views, the combined measure distinguished standard from nonstandard images with an area under the curve above 0.90. Performance was strongest for the apical four-chamber view, while some short-axis views were more challenging. Accuracy remained above 0.80 across the evaluated views.
The score also tracked gradual changes in image quality. In a separate internal dataset containing several cardiac disease groups, uncertainty declined as images moved from poor towards excellent quality. This pattern appeared across all assessed views.
Structure-level results showed which anatomical features became less visible as the probe moved away from the correct position. In parasternal views, changes commonly affected valves and the left ventricle. In apical views, chamber visibility was more informative. The structure-based measure therefore contributed both to overall scoring and to identifying the anatomical source of deterioration during acquisition in continuous image sequences. The same trend appeared as standard views progressively degraded.
Real-Time Guidance Requires Broader Validation
The probe-movement predictor achieved its strongest result in the apical four-chamber view, reaching 85.1% accuracy. Performance was lower in two views that require more precise positioning and are affected by greater anatomical or motion-related complexity. The hold instruction performed consistently well, while the results varied more for individual tilt, rock and rotation movements.
Reusing anatomical and classification features also improved data efficiency. Compared with a model trained only on raw ultrasound images, the shared-feature approach reached strong performance with half of the available training data. During continuous sequences, the predicted movement changed as image quality deteriorated and the uncertainty of disappearing structures increased.
The framework processed each frame in 17 milliseconds on one of the tested graphics processors, supporting real-time operation. The measurement included image preparation and calculation of the quality scores but did not include direct video transfer from the ultrasound device.
Further validation is required before clinical use. The guidance dataset was relatively small and consisted of adults without known heart disease. Performance needs assessment across diseases, populations and emergency settings. The current system also assigns one discrete action to each frame, although scanning may require combined and continuous movements. Errors in classification or segmentation can affect later outputs, while image noise, variable echogenicity and motion artefacts remain difficult to capture fully. The framework therefore remains a technical validation rather than a completed clinical tool.
The framework integrates quality assessment and corrective probe guidance within a single lightweight network. It distinguishes standard from degraded views, identifies structures associated with poorer visualisation and predicts practical probe movements without requiring separate quality labels. Its real-time speed and reduced dependence on manually annotated data support further development for training and portable echocardiography. However, the findings come from a limited guidance dataset and performance varies between imaging views. Broader clinical validation, video-based analysis and guidance for continuous movements are needed before the approach can support routine echocardiographic acquisition.
Channel: IM Community: AI
Source: IEEE Journal of Biomedical and Health Informatics
Image Credit: iStock
References:
Jeong H et al (2026) Multi-Task Deep Learning Framework for Real-Time Quality Assessment and Probe Guidance in Echocardiography. IEEE Journal of Biomedical and Health Informatics. doi: 10.1109/JBHI.2026.3717967.
