Ultrasound-based thyroid nodule segmentation with deep hybrid convolutional network

Automatic segmentation of ultrasound-based thyroid nodules can assist physicians in more efficiently and accurately assessing thyroid diseases. However, thyroid ultrasound imaging presents certain unique challenges, such as blurred boundaries of nodules, strong heterogeneity in tissue echoes, and limited contrast between the lesions and adjacent normal thyroid tissue, which hinder the precise identification of nodules in clinical assessment.

To overcome these obstacles, we propose a deep hybrid convolutional network, named UTNseg, for precisely automatic segmentation of ultrasound-based thyroid nodules.

UTNseg contains four newly designed modules upon TransUNet: (1) the spatial-channel feature strong capture (STransformer) module is designed to reinforce the ability of encoder on capturing spatial and channel features, (2) the redundant feature map utilization (RMU) module is devised to mitigate redundant information in the skip connections, (3) the composite convolution (CConv) module is newly designed to further preserve the key multi-scale features of the target thyroid nodules, (4) the lightweight edge refinement (LER) module is newly proposed to refine the initial segmentation outputs. UTNseg was evaluated on two publicly available ultrasound datasets of thyroid nodules (TN3K with 3493 images and DDTI with 637 images). Datasets were divided at the patient level, with 4000 training images (2000 images from TN3K and 2000 images from DDTI, augmented by random rotation, scaling, and horizontal flipping on 445 images), 761 validation images (698 images from TN3K and 63 images from DDTI), and 924 test images (795 images from TN3K and 129 images from DDTI). Additionally, separate experiments were performed on these two datasets in comparative experiments. And the performance was compared with six state-of-the-art segmentation methods (U-Net, TransUNet, TRFENet, UNeXt, TRFE+ and BPAT-UNet). Evaluation metrics mainly included Dice similarity coefficient (DSC) and 95th percentile Hausdorff distance (HD95). Statistical significance level was analyzed through the Wilcoxon signed-rank test, and the effect size was quantified by Cohen's $d$ . Holm-Bonferroni correction was employed to correct for multiple comparisons.

On the TN3K dataset, UTNseg achieved 79.52% DSC and 16.32 HD95. On the DDTI dataset, UTNseg achieved 86.21% DSC and 12.76 HD95. On the hybrid dataset, UTNseg achieved 80.77% DSC and 14.56 HD95. These results showed that UTNseg achieved superior performance. Compared with state-of-the-art approaches, UTNseg was statistically significant ( $p<0.05$ ), and the corresponding effect sizes were mostly medium or large, further quantifying the practical importance of the improvements.

UTNseg demonstrated excellent segmentation performance on the public dataset. Statistical significance and effect size analysis both indicate that the improvements are reliable and have practical significance.