SAM-Driven Weakly Supervised Nodule Segmentation with Uncertainty-Aware Cross Teaching

TL;DR

This work tackles the high cost of pixel-level annotations in ultrasound nodule segmentation by leveraging the segmentation foundation model SAM to generate pseudo-labels from aspect ratio annotations. It introduces a two-stage approach: (i) construct shape-prior bounding-box prompts (tight box, ellipse, and minimum enclosing circle) to elicit diverse SAM predictions and (ii) select and refine pseudo-labels, followed by a SAM-driven uncertainty-aware cross-teaching scheme with two networks trained on complementary pseudo-labels and supervised within uncertain regions. The method defines an uncertainty map $U_i = Y_{i}^{int} oxor Y_{i}^{uni}$ and optimizes a total loss $L_{total} = L_{sup} + lambda L_{ct-u}$ to mitigate label noise. Experiments on thyroid and breast ultrasound datasets show state-of-the-art performance among weakly supervised methods, reducing annotation burden while maintaining high segmentation quality, with potential generalization to other medical imaging tasks.

Abstract

Automated nodule segmentation is essential for computer-assisted diagnosis in ultrasound images. Nevertheless, most existing methods depend on precise pixel-level annotations by medical professionals, a process that is both costly and labor-intensive. Recently, segmentation foundation models like SAM have shown impressive generalizability on natural images, suggesting their potential as pseudo-labelers. However, accurate prompts remain crucial for their success in medical images. In this work, we devise a novel weakly supervised framework that effectively utilizes the segmentation foundation model to generate pseudo-labels from aspect ration annotations for automatic nodule segmentation. Specifically, we develop three types of bounding box prompts based on scalable shape priors, followed by an adaptive pseudo-label selection module to fully exploit the prediction capabilities of the foundation model for nodules. We also present a SAM-driven uncertainty-aware cross-teaching strategy. This approach integrates SAM-based uncertainty estimation and label-space perturbations into cross-teaching to mitigate the impact of pseudo-label inaccuracies on model training. Extensive experiments on two clinically collected ultrasound datasets demonstrate the superior performance of our proposed method.

Figures (3)

• Figure 1: Examples for aspect ratio annotation and prediction results of segment anything model in different bounding box prompts setting.
• Figure 2: Overview of our proposed framework. The framework includes two stages. In the initial stage, three distinct sets of prior bounding boxes are created based on clinical annotations and fed as prompts into the SAM, generating pseudo labels. In the subsequent stage, a selection is made from these pseudo labels. A strategy named SAM-driven Uncertainty-Aware Cross Teaching is then introduced, leveraging these selected pseudo labels for nodule segmentation.
• Figure 3: The impact of varying $\lambda$ on the segmentation performance of the proposed framework, assessed on the Thyroid Ultrasound dataset using Dice similarity coefficient (DSC) and 95th percentile Hausdorff Distance (HD95).