Segmentation-Based vs. Regression-Based Biomarker Estimation: A Case Study of Fetus Head Circumference Assessment...

Area of Science:

• Medical imaging
• Artificial intelligence in healthcare
• Fetal biometry

Background:

• Fetal head circumference (HC) is crucial for monitoring fetal growth during pregnancy using ultrasound (US) images.
• Current automated HC estimation relies on segmentation networks followed by ellipse fitting, with segmentation as an intermediate step.

Purpose of the Study:

• To quantitatively compare segmentation-based and segmentation-free (regression) deep learning approaches for fetal HC estimation.
• To assess the performance gap between direct regression and indirect segmentation-based methods.

Main Methods:

• Experimented with various convolutional neural network (CNN) architectures and backbones for both segmentation and regression models.
• Conducted fair, quantitative comparisons on the HC18 dataset, including agreement analysis.
• Investigated memory usage and computational efficiency for both approaches.

Main Results:

• Segmentation-based approaches yielded the most accurate HC estimations.
• Regression CNN approaches demonstrated learning of prominent features, offering promising but improvable HC estimation results.
• Both methods were compared for accuracy, memory usage, and computational efficiency.

Conclusions:

• While segmentation-based methods currently offer superior accuracy for fetal HC estimation, direct regression approaches show potential.
• Further research and development are needed to improve the accuracy of segmentation-free regression methods for fetal biometry.