Ultrasound Nodule Segmentation Using Asymmetric Learning with Simple Clinical Annotation

TL;DR

This paper tackles the high cost of precise annotations in ultrasound nodule segmentation by exploiting readily available clinical aspect ratio annotations. It proposes an asymmetric two-branch framework (CRBNet) that converts annotations into conservative and radical pseudo-labels, and further integrates them via Conservative-Radical-Balance Strategy (CRBS), Inconsistency-Aware Dynamically Mixed Pseudo Labels Supervision (IDMPS), and a Clinical Anatomy Prior Loss (CAP) to leverage spatial priors. Through experiments on thyroid and breast ultrasound datasets, the method achieves state-of-the-art performance among weakly supervised approaches and even matches or exceeds fully supervised results, while substantially reducing annotation effort. The work also provides a new ultrasound nodule dataset with both aspect-ratio and ground-truth segmentations, and outlines extensions to 3D segmentation, highlighting practical impact for clinical deployment and future research directions.

Abstract

Recent advances in deep learning have greatly facilitated the automated segmentation of ultrasound images, which is essential for nodule morphological analysis. Nevertheless, most existing methods depend on extensive and precise annotations by domain experts, which are labor-intensive and time-consuming. In this study, we suggest using simple aspect ratio annotations directly from ultrasound clinical diagnoses for automated nodule segmentation. Especially, an asymmetric learning framework is developed by extending the aspect ratio annotations with two types of pseudo labels, i.e., conservative labels and radical labels, to train two asymmetric segmentation networks simultaneously. Subsequently, a conservative-radical-balance strategy (CRBS) strategy is proposed to complementally combine radical and conservative labels. An inconsistency-aware dynamically mixed pseudo-labels supervision (IDMPS) module is introduced to address the challenges of over-segmentation and under-segmentation caused by the two types of labels. To further leverage the spatial prior knowledge provided by clinical annotations, we also present a novel loss function namely the clinical anatomy prior loss. Extensive experiments on two clinically collected ultrasound datasets (thyroid and breast) demonstrate the superior performance of our proposed method, which can achieve comparable and even better performance than fully supervised methods using ground truth annotations.

Figures (6)

• Figure 1: Visualization of clinical aspect ratio annotation and the corresponding pseudo labels. The pink lines depict the edges of the ground truth, the orange lines represent the edges of the radical labels, the blue lines indicate the edges of the conservative labels, and the yellow lines display the predicted edges generated by the model trained using the respective pseudo label.
• Figure 2: The overall framework of our proposed method. The areas highlighted with a gray background represent the CRB strategy, while those with a blue background correspond to the IDMPS module. The conservative model and conservative label are indicated in blue, whereas the radical model and radical label are depicted in orange.
• Figure 3: Performance sensitivity to hyper-parameter $\alpha$ and loss weights, $\lambda_{1}$ and $\lambda_{2}$ on the thyroid ultrasound dataset.
• Figure 4: The visual segmentation examples of different comparison methods and the ground truth. The yellow curves show ground truth segmentation.
• Figure 5: Quantitative results of IDMPS with various combinations of conservative and radical labels. We measure the corresponding DSC on the breast ultrasound dataset.