To assess neural network performance in ultrasound image segmentation, emphasizing the importance of distinguishing between distinct and completed anatomical borders for improved evaluation.
Key Findings:
Neural networks are typically evaluated by comparing predictions to expert labels, which may not reflect their understanding of ambiguous borders, impacting diagnostic accuracy.
Grad-CAM is effective for analyzing network activations, yet qualitative evaluations of ultrasound borders are lacking, which is crucial for clinical applications.
The LEG-3D-US dataset presents unique challenges due to non-homogeneous tissues and unclear borders, necessitating advanced segmentation techniques.
Interpretation:
The study emphasizes the need for a nuanced evaluation of segmentation networks in ultrasound imaging, focusing on both clear and ambiguous anatomical boundaries, which could enhance diagnostic accuracy and treatment planning.
Limitations:
The study primarily focuses on four ultrasound datasets, which may limit generalizability; future work should explore additional datasets.
Existing methods for model interpretation may not fully capture the complexities of ultrasound imaging, suggesting a need for developing more tailored interpretative frameworks.
Conclusion:
This work contributes to the understanding of ultrasound segmentation networks by providing a novel evaluation framework and an open-source dataset, promoting further research in the field.
by Vanessa Gonzalez Duque, Alexandra Marquardt, Yordanka Velikova, Lilian Lacourpaille, Antoine Nordez, Marion Crouzier, Hong Joo Lee, Diana Mateus, Nassir Navab
