Wavelet‐Based Frequency Replacement and Edge Enhancement for Semi‐Supervised Fetal Ultrasound Segmentation: A Methodological Review and Technical Analysis

We read with great interest the article by Yue et al, “Wavelet-Based Frequency Replacement and Edge Enhancement for Semi-Supervised Fetal Ultrasound Image Segmentation,”1 published in the Journal of Ultrasound in Medicine. The authors propose a semi-supervised framework integrating discrete wavelet transform (DWT) with edge enhancement to address low contrast and blurred boundaries in fetal ultrasound. By introducing a weak-to-strong frequency component replacement (WSCR) strategy and an edge-aware gradient enhancement (EAGE) module, their method achieves significant segmentation improvements using minimal annotated data, offering a promising direction for reducing annotation burden in clinical tasks such as fetal biometry. This study has several notable strengths. First, the frequency-domain augmentation via DWT preserves global structure while introducing beneficial high-frequency perturbations, surpassing traditional spatial augmentations. Second, the EAGE module guides model attention toward indistinct boundaries through wavelet-derived gradient masks. Third, comprehensive validation on 3 public datasets (PSFHS, HC18, CCAUI) demonstrates robustness, and ablation studies confirm the synergistic contribution of each component. Despite these contributions, 1 methodological aspect warrants further discussion. The heuristic selection of which high-frequency components (LH, HL, HH) to replace, while effective, may not be universally optimal. Recent work on adaptive feature selection2 and the CR-SCAD algorithm—which balances dense representation with sparse variable selection3—suggests that learning-based weighting of wavelet sub-bands could enhance flexibility across varying edge orientations. Incorporating such adaptive mechanisms could further improve segmentation in diverse clinical scenarios. Relatedly, aggregation of multi-scale frequency information could be refined. The current concatenation and binary masking may not fully capture interactions among anatomical structures. Contrast- and gain-aware attention4 offers adaptive fusion that preserves dominant features. Extending this concept to WSCR could dynamically emphasize informative sub-bands (eg, diagonal for curved head contours, horizontal for linear symphysis), better preserving structural characteristics while suppressing noise. Clinically, enhanced boundary detection promises to reduce inter-operator variability in measurements such as fetal head circumference, improving gestational age assessment. However, the reported preprocessing time (0.82 s/image) remains below real-time requirements (10–100 fps). Future work should prioritize computational efficiency through end-to-end learnable wavelet layers. Additionally, leveraging rich datasets such as PSFHS5 for prospective validation will be essential to translate these technical gains into clinical practice. In conclusion, Yue et al provide a valuable contribution by demonstrating the power of frequency-domain augmentation and edge enhancement in semi-supervised ultrasound segmentation. Future exploration of adaptive frequency selection,2, 3 learnable feature aggregation,4 and rigorous clinical validation5 will further strengthen the impact of this approach. We commend the authors for their rigorous work and hope these observations encourage continued refinement. Data sharing not applicable to this article as no datasets were generated or analyzed during the current study.