Human-level Performance On Automatic Head Biometrics In Fetal Ultrasound Using Fully Convolutional Neural Networks (1804.09102v1)

Abstract: Measurement of head biometrics from fetal ultrasonography images is of key importance in monitoring the healthy development of fetuses. However, the accurate measurement of relevant anatomical structures is subject to large inter-observer variability in the clinic. To address this issue, an automated method utilizing Fully Convolutional Networks (FCN) is proposed to determine measurements of fetal head circumference (HC) and biparietal diameter (BPD). An FCN was trained on approximately 2000 2D ultrasound images of the head with annotations provided by 45 different sonographers during routine screening examinations to perform semantic segmentation of the head. An ellipse is fitted to the resulting segmentation contours to mimic the annotation typically produced by a sonographer. The model's performance was compared with inter-observer variability, where two experts manually annotated 100 test images. Mean absolute model-expert error was slightly better than inter-observer error for HC (1.99mm vs 2.16mm), and comparable for BPD (0.61mm vs 0.59mm), as well as Dice coefficient (0.980 vs 0.980). Our results demonstrate that the model performs at a level similar to a human expert, and learns to produce accurate predictions from a large dataset annotated by many sonographers. Additionally, measurements are generated in near real-time at 15fps on a GPU, which could speed up clinical workflow for both skilled and trainee sonographers.

• The paper demonstrates that using FCN-based segmentation achieves measurement errors close to expert performance (HC error of 1.99 mm vs. 2.16 mm and BPD error of 0.61 mm vs. 0.59 mm).
• The methodology employs a pre-trained FCN with semantic segmentation and ellipse fitting on a large dataset, effectively mimicking clinical measurement techniques.
• The approach operates near real-time at 15 frames per second, offering practical improvements in clinical workflow and consistency in fetal ultrasound examinations.

Human-level Performance On Automatic Head Biometrics In Fetal Ultrasound Using Fully Convolutional Neural Networks

The paper presents a method leveraging Fully Convolutional Networks (FCNs) for achieving human-level performance in the measurement of fetal head biometrics such as head circumference (HC) and biparietal diameter (BPD) from ultrasound images. The primary motivation is to reduce the significant inter-observer variability that often occurs in clinical settings due to the skill level required and the variability of ultrasound imaging conditions.

Methodology

The methodology centers around utilizing a FCN trained on a substantial set of ultrasound images annotated by a large cohort of sonographers. Approximately 2000 images were used, which is a relatively large dataset for this type of medical imaging task. An FCN, initialized with pre-trained ImageNet weights, was adapted to the task through a semantic segmentation approach. The segmentation of the head is achieved with the trained FCN, after which an ellipse is fitted to the contours of the segmentation to obtain biometric measurements. This approach mimics standard clinical practices but automates the task with comparable accuracy to human experts.

Experimental Design and Results

The authors focused on evaluating the proposed method's accuracy against human experts by comparing the model-derived biometric measurements with those manually annotated by experts on a test set of 100 images. Additionally, they provided an error analysis using mean absolute error (MAE) and mean error (ME) metrics, along with Dice coefficients for segmentation accuracy.

Key findings include:

• The mean absolute model-expert error for HC was 1.99 mm, slightly outperforming the inter-expert error of 2.16 mm.
• For BPD, the model-expert error was 0.61 mm, effectively equivalent to the inter-expert error of 0.59 mm.
• Further, the proposed method demonstrated a Dice coefficient of 0.980, consistent with expert-level performance.
• Importantly, the method operates in near real-time at 15 frames per second, potentially enhancing clinical workflow efficiency.

Implications

The implications of this work are substantial for clinical practice. With its high accuracy and real-time capabilities, the FCN-based approach could be used to support sonographers, especially less experienced ones, in achieving consistent and reliable measurements. This could be particularly advantageous in regions with varying availabilities of skilled clinical personnel. The automation of fetal head biometric measurement also promises to reduce the time required per examination and improve the clinical workflow.

Additionally, from a research perspective, this work can be extended further to measure other biometric dimensions like abdominal circumference and femur length, addressing the challenges of prenatal screening comprehensively. The methodological framework presented can also adapt to various imaging systems beyond the specific ultrasound devices used in the paper, allowing for broader applicability through techniques like domain adaptation or fine-tuning.

Conclusion

This research demonstrates a significant advancement in automated ultrasound biometrics. The work suggests future directions in AI-assisted diagnostic tools in medical imaging, offering robust solutions to common challenges faced in the field. Moreover, the dissemination of such models can empower clinicians worldwide, potentially standardizing the quality of prenatal care. The results reflect a successful integration of machine learning into practical medical applications, setting a foundation for ongoing innovations in this domain.