A weakly supervised deep active contour model for nodule segmentation in thyroid ultrasound images

Nodule segmentation is crucial for thyroid ultrasound image analysis. Recent progress in this task is driven by deep learning methods and large high-quality annotated datasets. However, obtaining pixel-wise labels are labor-intensive and time-consuming. Weakly supervised learning is proposed to address limited annotations with simple labels, but most are not satisfactory in capturing blurred boundaries due to the lack of attention to edges. In this paper, we propose a novel weakly supervised deep active contour model for nodule segmentation in thyroid ultrasound images. The key idea is to deform an initial contour iteratively to match thyroid nodule boundary by regressing vertex offsets which are obtained by gradient similarity and statistical information. First, a polygon contour is adopted as initial label, which aims to re-duce interference from surrounding organs. Second, contour deformation network deforms initial contour by regressing vertex-wise offsets. We introduce a loss function based on level-set theory and an auxiliary edge attention module to enhance ability to capture fuzzy boundaries. As demonstrated by experimental results, our model achieved a dice coefficient of 0.87 +/- 0.07 and Harsdorf distance of 17.11 +/- 7.95, outper-forming other state-of-the-art weakly supervised methods. The segmentation results are effective with fully supervised methods, but our work lighten annotation burdens. Therefore, our algorithm can be ef-fectively used in clinical diagnosis.(c) 2022 Elsevier B.V. All rights reserved.

Automated summary

This paper proposes a novel weakly supervised deep active contour model for nodule segmentation in thyroid ultrasound images. The model achieves accurate segmentation results by iteratively deforming an initial contour to match the nodule boundary.