A preliminary study of a hybrid transformer with enhanced local feature fusion for nuchal translucency ultrasound image segmentation.

BACKGROUND

Nuchal translucency (NT) is an important ultrasound indicator in early pregnancy. Increased NT is strongly associated with chromosomal abnormalities, structural defects, and other adverse outcomes. However, the low signal-to-noise ratio (SNR), numerous artifacts, and poor contrast present in ultrasound images pose considerable challenges to existing segmentation methods.

PURPOSE: To address these challenges, this paper proposes a novel Transformer model, TransNTSeg, for NT segmentation. This study serves as a preliminary proof-of-concept to evaluate the feasibility of the proposed architecture.

METHODS

A clinical dataset of 123 B-mode ultrasound images is collected. To ensure rigorous evaluation, the dataset is partitioned into 98 images for training/validation (using 5-fold cross-validation) and 25 images for independent testing. The architecture incorporates two main innovations: first, an enhanced Dense Dilated Mixed Feed-Forward Network (DDMix-FFN) module within the Transformer that more robustly integrates global dependencies and local context to handle image noise and boundary ambiguity; and second, a DDMix-FFN Transformer context bridge that fuses multi-scale encoder features for precise localization of the small NT region. Performance is compared against eight state-of-the-art benchmarks (e.g., U-Net, MT-UNet) using Intersection over Union (IoU) and Dice Similarity Coefficient (DSC). Statistical significance is assessed using the paired t-test with Bonferroni correction to control for multiple comparisons (adjusted p < 0.0015625), and practical significance is quantified using Cohen's dz effect size.

RESULTS

TransNTSeg achieves the highest performance on the test dataset, with a mean IoU of 0.763 ± 0.05 and a mean DSC of 0.852 ± 0.03. Rigorous statistical analysis (Bonferroni corrected p < 0.0015625) confirms considerable improvements. Based on the evaluation of effect sizes, TransNTSeg is identified as the best-performing method, demonstrating a substantial effect size (d = 0.745) against the closest competitor (MT-UNet).

CONCLUSION

While demonstrating competitive performance within this controlled setting, these promising outcomes highlight the potential of TransNTSeg for accurate NT segmentation. This warrants further validation on multicenter, multi-vendor datasets to establish its generalizability for broad clinical use.