TL;DR
This paper introduces and clinically evaluates a deep learning (DL) algorithm for automatically extracting standard 2D fetal head planes from 3D ultrasound volumes. The DL approach shows comparable quality to operator-selected planes, suggesting a path towards more accessible and efficient 3D ultrasound in clinical practice, though overall 3D image quality still lags behind traditional 2D methods.
What Problem Does This Paper Address?
Routine fetal anomaly scans rely on specific 2D ultrasound (US) views, like the transventricular (TV) and transcerebellar (TC) planes, to assess fetal development and identify potential anomalies. While 3D US offers the potential for more detailed retrospective analysis and easier manipulation of image planes, manually selecting and extracting these standard planes from 3D volumes is time-consuming and requires significant expertise. This creates a barrier to widespread clinical adoption of 3D US for routine screening. This paper addresses this by developing and evaluating an automated method to extract these crucial planes, aiming to bridge the gap between the potential of 3D US and its practical clinical application.
Methods
The core of this work is a deep learning (DL) algorithm based on an iterative transformation network (ITN) approach. This algorithm learns to map a 2D image to the required transformation to align it with a standard plane within a 3D volume. The ITN model was trained using manually annotated fetal head landmarks from 3D US volumes. For clinical evaluation, two observers retrospectively assessed the quality of automatically extracted planes (3D-DL) against manually acquired 2D planes (2D-MA) and operator-selected 3D planes (3D-OS). A standardized quality assessment tool (USQA) and pre-defined criteria were used to evaluate image adequacy, focusing on visualization of key anatomical landmarks. Inter- and intra-observer agreement was also assessed.
Datasets
The study utilized 3D fetal head ultrasound volumes acquired from 91 transabdominal scans. For algorithm training, 303 TV and 248 TC planes were manually annotated. For the clinical evaluation phase, a subset of 91 TV and 73 TC images were selected. These images were then compared across three modalities: 2D manually-acquired (2D-MA), 3D operator-selected (3D-OS), and 3D DL automatically-extracted (3D-DL).
Key Insights and Ideas
- DL as a viable alternative to manual selection: The study demonstrates that a DL algorithm can automatically extract standard fetal head planes with a quality comparable to planes manually selected by experienced operators from the same 3D volume. This is a significant step towards automating a previously labor-intensive process.
- 3D US quality limitations persist: Despite the success of the automated extraction, the overall image quality of the 3D volumes themselves (and thus the extracted planes) was found to be inferior to conventional 2D planes. This suggests that current 3D acquisition techniques and the inherent challenges of fetal imaging (e.g., ossification, fetal positioning) are the primary bottlenecks for achieving high-quality 3D fetal neurosonography.
- GA matters for plane visibility: The study hints that lower gestational ages (GA) might be more amenable to accurate plane localization by automated tools, likely due to better visibility of intracranial structures before significant ossification.
Key Results
- The DL algorithm achieved a detection accuracy of approximately 3.7 mm and 12.7° for TV and TC planes, respectively, indicating good localization.
- For TV planes, 67% of 3D-OS and 3D-DL planes were deemed adequate quality, while for TC planes, only 45% of 3D-OS and 55% of 3D-DL planes met adequacy criteria.
- A significant difference in image quality was observed between 2D-MA and 3D modalities (both OS and DL), with 2D planes generally superior. However, no significant difference in quality was found between the 3D operator-selected and 3D DL automatically-extracted planes.
- Inter-observer agreement for overall adequacy was moderate for TV planes (Kappa ~0.6) and weak to non-existent for TC planes, highlighting the subjective nature of quality assessment, especially for TC views.
Limitations and Future Work
The authors acknowledge that the quality of the 3D volumes themselves remains a limitation, impacting the quality of extracted planes. This is attributed to factors like fetal head position, ossification artifacts, and the limitations of the ultrasound transducer. The study also notes that the wider GA range of the dataset, due to research project requirements, might have influenced the results, suggesting that future work should focus on refining the algorithm using datasets from earlier gestations. The clinical value of these automated tools still needs to be definitively established. Future work could also explore the automatic extraction of sagittal and coronal planes to provide a more complete fetal neurosonographic examination.
Summary updated