Report

Teaser, summary, work performed and final results

Periodic Reporting for period 1 - MOUSIE (Multi-Organ UltraSound-based Inborn Evaluation)

Teaser

Two-dimensional (2D) ultrasound (US) screening is the preferred imaging modality for prenatal evaluation for growth, gestational age estimation, and early structural abnormalities detection. However, prenatal screening still relies on 2D measurements of different organs that...

Summary

Two-dimensional (2D) ultrasound (US) screening is the preferred imaging modality for prenatal evaluation for growth, gestational age estimation, and early structural abnormalities detection. However, prenatal screening still relies on 2D measurements of different organs that are analysed separately from organ-specific standard planes, manually acquired using a hand-held prove, thus suffering from subjectivity, low reproducibility, and operator dependency.
The main purpose of the MOUSIE project was to develop new image-based machine learning solutions for a more efficient, quantitative and accurate prenatal healthcare. Thanks to the recent progress in deep learning techniques, it is now possible to conceive image analysis frameworks that operate on US data alone, able to analyse the rich and dense information of this modality in a more efficient and accurate way. The shift of the machine learning community towards these deep learning-based solutions, and the great potential of this technology in the medical image analysis field, were early identified in the project, thus adapting the research plan accordingly.

The work carried out in this project has revolved around two main fundamental pillars: i) the development of new machine learning solutions for the efficient and accurate analysis of fetal sonographic images; and ii) the use of 3D US volumetric data to mitigate the inherent practical and diagnostic limitations of traditional 2D sonographic scans.

Work performed

The objectives of the MOUSIE project were early revised and adapted in light of the recent progress in the field of machine learning.
More specifically, the objectives of the tasks accomplished in the project are:

• Development of new machine learning frameworks for the automatic segmentation of the fetal skull in 3D US volumes. 3D US has the capacity to mitigate many of the limitations of 2D scans, reducing subjectivity, and improving diagnosis accuracy and reproducibility. However, despite the potential of volumetric data, it is necessary to develop new technological solutions for the efficient analysis, visualization and segmentation of the clinically relevant information in 3DUS. In this line, I have created different segmentation architectures able to automatically delineate the fetal skull from 3DUS volumes.

• Development of new frameworks for optimal volumetric data acquisition, visualization and navigation. Traditionally, sonographers have been trained to acquire optimal 2D planar views with the free-and US probe. However, the acquisition of new 3D data may be affected by the lack of experience with this new technology, reducing the quality of the data, or slowing down the acquisition process as compared to traditional 2D US. Additionally, 3D US imaging also poses new challenges, such as the efficient visualization and navigation of dense volumetric data. The creation of new tools to address these problems is key in for the successful integration of 3D sonography into the clinical practice.

• Creation of a new generation of 3D biometrics for the assessment of fetal head dismorphology. Traditionally, the assessment of fetal skull deformations was based on simple 2D biometrics manually extracted from single 2D planar views of the head. Due to this oversimplified approach, the early identification of potential structural anomalies is highly subjective, and operator-dependent, with the detection rates remaining below the recommended values. Thanks to the computational power provided by recent machine learning models, it is now possible to extract rich 3D anatomical information directly from 3D US volumes, allowing for a more accurate, robust, and objective analysis of the fetal anatomy.

• Development of new generative models for the 3D reconstruction of the fetal anatomy from 2D views. Despite the potential of 3D fetal biometry, its clinical impact can be notably hampered by two important factors: i) the lack of a significant number of cases to create statistically relevant anatomical models; ii) the limited access to 3D US transducers, especially in developing countries and remote areas. Aware of these limitations, another important part of the project was to investigate new technological solutions that contribute to the popularization of 3D-based fetal anatomical analysis, allowing for large-scale 3D-based biometric studies that include a wide and varied demographic representation. In this context, the project also investigated new computational solutions for the 3D reconstruction of the fetal skull from 2D US standard planes routinely acquired in the clinical practice.

Final results

Regarding its social impact, all the technological solutions developed during the MOUSIE project have a direct clinical translation, contributing to the progress towards a more informed, quantitative, and objective practice of prenatal healthcare. In particular, the MOUSIE project has pioneered some of the most advanced machine learning frameworks for the use of 3D US imaging as an integral part of the mid-trimester morphology scan. These tools were designed for an efficient and natural integration into the clinical workflow, allowing for an accurate analysis of the fetal anatomy, and minimizing the inherent subjectivity and operator dependency of traditional 2D US scans. Moreover, the MOUSIE project also introduced a new generation of 3D fetal biometrics, proving superior diagnostic capabilities and robustness than conventional manually defined 2D metrics. The solutions developed during this project are being integrated as part of a clinical pilot study on advanced fetal imaging, in collaboration with King’s College London and St. Thomas Hospital London (visit http://www.ifindproject.com/ for more details about this collaborative project). During this study, the tools will be tested in a clinical environment and with real patients as part of the second trimester sonographic scan, providing valuable feedback from patients and the clinical team. Finally, it is also important to highlight the potential impact of these new technological solutions to improve healthcare in particularly challenging regions, including developing countries, and remote areas. The limited availability of specialized clinical experts and resources results in a suboptimal prenatal healthcare, presenting higher rates of mortality and lower abnormality detection rates. Automatic machine learning-based solutions will help to enhance clinical performance, reducing diagnostic errors and inaccurate subjective evaluations.

To facilitate the widespread diffusion of the solutions proposed, all the published papers are freely available from the researcher website (http://juanjosecerrolaza.com/publications/), where additional details of the project (http://juanjosecerrolaza.com/research/), and related news (http://juanjosecerrolaza.com/news/), are also provided. Despite the code is not ready for open access yet, detailed instructions for the implementation of the solutions proposed are provided in the papers.

Website & more info

More info: http://juanjosecerrolaza.com/.