Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Improving Automatic Fetal Biometry Measurement with Swoosh Activation Function

Shijia Zhou, Euijoon Ahn, Hao Wang, Ann Quinton, Narelle Kennedy, Pradeeba Sridar, Ralph Nanan, Jinman Kim

TL;DR

This work tackles the challenge of accurately measuring fetal biometrics (FTD and FHC) from noisy 2D ultrasound by addressing the limitations of the state-of-the-art BiometryNet in handling fuzzy edges and complex thalamus geometry. The authors introduce the Swoosh Activation Function (SAF), a regularization mechanism that shapes heatmaps by enforcing an optimal MSE between predicted landmark heatmaps and ground-truth Gaussian heatmaps, while also constraining each heatmap against a zero baseline. SAF is architecture-agnostic and configurable, demonstrated to improve FTD ICC and FHC mean differences across BiometryNet and EfficientNet on two public datasets (FTD and HC18). The approach is simple to implement, shows clear performance gains, and holds promise for broader adoption in landmark-detection problems across ultrasound and other imaging modalities, potentially supporting better fetal monitoring and neonatal outcomes. The key contribution is a novel, tunable activation-based regularization that stabilizes heatmap predictions and enhances measurement accuracy without altering network architectures.

Abstract

The measurement of fetal thalamus diameter (FTD) and fetal head circumference (FHC) are crucial in identifying abnormal fetal thalamus development as it may lead to certain neuropsychiatric disorders in later life. However, manual measurements from 2D-US images are laborious, prone to high inter-observer variability, and complicated by the high signal-to-noise ratio nature of the images. Deep learning-based landmark detection approaches have shown promise in measuring biometrics from US images, but the current state-of-the-art (SOTA) algorithm, BiometryNet, is inadequate for FTD and FHC measurement due to its inability to account for the fuzzy edges of these structures and the complex shape of the FTD structure. To address these inadequacies, we propose a novel Swoosh Activation Function (SAF) designed to enhance the regularization of heatmaps produced by landmark detection algorithms. Our SAF serves as a regularization term to enforce an optimum mean squared error (MSE) level between predicted heatmaps, reducing the dispersiveness of hotspots in predicted heatmaps. Our experimental results demonstrate that SAF significantly improves the measurement performances of FTD and FHC with higher intraclass correlation coefficient scores in FTD and lower mean difference scores in FHC measurement than those of the current SOTA algorithm BiometryNet. Moreover, our proposed SAF is highly generalizable and architecture-agnostic. The SAF's coefficients can be configured for different tasks, making it highly customizable. Our study demonstrates that the SAF activation function is a novel method that can improve measurement accuracy in fetal biometry landmark detection. This improvement has the potential to contribute to better fetal monitoring and improved neonatal outcomes.

Improving Automatic Fetal Biometry Measurement with Swoosh Activation Function

TL;DR

This work tackles the challenge of accurately measuring fetal biometrics (FTD and FHC) from noisy 2D ultrasound by addressing the limitations of the state-of-the-art BiometryNet in handling fuzzy edges and complex thalamus geometry. The authors introduce the Swoosh Activation Function (SAF), a regularization mechanism that shapes heatmaps by enforcing an optimal MSE between predicted landmark heatmaps and ground-truth Gaussian heatmaps, while also constraining each heatmap against a zero baseline. SAF is architecture-agnostic and configurable, demonstrated to improve FTD ICC and FHC mean differences across BiometryNet and EfficientNet on two public datasets (FTD and HC18). The approach is simple to implement, shows clear performance gains, and holds promise for broader adoption in landmark-detection problems across ultrasound and other imaging modalities, potentially supporting better fetal monitoring and neonatal outcomes. The key contribution is a novel, tunable activation-based regularization that stabilizes heatmap predictions and enhances measurement accuracy without altering network architectures.

Abstract

The measurement of fetal thalamus diameter (FTD) and fetal head circumference (FHC) are crucial in identifying abnormal fetal thalamus development as it may lead to certain neuropsychiatric disorders in later life. However, manual measurements from 2D-US images are laborious, prone to high inter-observer variability, and complicated by the high signal-to-noise ratio nature of the images. Deep learning-based landmark detection approaches have shown promise in measuring biometrics from US images, but the current state-of-the-art (SOTA) algorithm, BiometryNet, is inadequate for FTD and FHC measurement due to its inability to account for the fuzzy edges of these structures and the complex shape of the FTD structure. To address these inadequacies, we propose a novel Swoosh Activation Function (SAF) designed to enhance the regularization of heatmaps produced by landmark detection algorithms. Our SAF serves as a regularization term to enforce an optimum mean squared error (MSE) level between predicted heatmaps, reducing the dispersiveness of hotspots in predicted heatmaps. Our experimental results demonstrate that SAF significantly improves the measurement performances of FTD and FHC with higher intraclass correlation coefficient scores in FTD and lower mean difference scores in FHC measurement than those of the current SOTA algorithm BiometryNet. Moreover, our proposed SAF is highly generalizable and architecture-agnostic. The SAF's coefficients can be configured for different tasks, making it highly customizable. Our study demonstrates that the SAF activation function is a novel method that can improve measurement accuracy in fetal biometry landmark detection. This improvement has the potential to contribute to better fetal monitoring and improved neonatal outcomes.

Paper Structure

This paper contains 14 sections, 2 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Method
  3. Swoosh Activation Function
  4. SAF Regularization
  5. Datasets
  6. FTD Dataset:
  7. HC18 Dataset:
  8. Experimental Setup
  9. Training Epochs and Learning Rate
  10. Pre-processing
  11. SAF Configuration
  12. Results
  13. Discussion
  14. Conclusion

Figures (4)

  • Figure 1: Left: 2D-US image of fetal head from the HC18 dataset ThomasL.A.vandenHeuvel2018AutomatedImages, note part of the skull is not mineralized and has similar echogenicty to the adjacent uterine tissue. Middle: 2D-US image of a fetal femur ApostolosKolitsidakis2021HowLength. Right: 2D-US image of a GsS for measuring FTD, note the gaps in the skull due to unfused bones.
  • Figure 2: A: a 2D-US fetal brain image that has the GsS (outlined with white curly lines) annotated with manually created bounding box (white box); B: input image of the GsS constrained by the bounding box. The red dots represent the ground truth landmarks of FTD; C: the first heatmap of one of the FTD landmarks with the hottest (red) spot representing the landmark; D: the second heatmap of the other FTD landmark with the hottest (red) spot representing the landmark
  • Figure 3: A (left column): a pair of ground truth heatmaps with optimum MSE = 0.0061. B (middle column): a pair of predicted heatmaps with low MSE = 0.00003. C (right column): a pair of predicted heatmaps with high MSE = 17.7487
  • Figure 4: Row A: landmarks predicted and heatmaps produced by BiometryNet. Left, input image overlaid with predicted landmarks (red spots). Middle, the first predicted heatmap, there are hotspots present near both upper and lower wing-tips. Right, the second predicted heatmap. Row B: landmarks predicted and heatmaps produced by BiometryNet_SAF_a1. Left input image overlaid with predicted landmarks (red spots). Middle, the first predicted heatmap. Right, the second predicted heatmap.