Box2Pseudo: A Semi-Supervised Learning Framework for Pulmonary Nodule Segmentation with Box-Prompt Pseudo Supervision

Accurate pulmonary nodule segmentation is critical for early diagnosis of lung cancer. Yet, the high cost and labor-intensive nature of pixel-wise manual annotations remains challenging. In real-world clinical practice, weakly annotated labels like bounding boxes are more affordable and they always coexist with fully pixel-level annotated labels. However, the simultaneous use of fully and weakly annotated data for pulmonary nodule segmentation presents complexities. In this paper, we propose Box2Pseudo, a principled semi-supervised framework for pulmonary segmentation that only uses a small set of fully labeled data (having pixel-level and box labels) and a large set of weakly labeled data (having box labels only). Specifically, our Box2Pseudo consists of three networks, including the box-prompt network (BPN), the pseudo-refine network (PRN) and the main network (MAN). To make full use of localization priors provided by bounding boxes, we propose background-filter layer (BFL), which can be combined with BPN and PRN to generate high-quality pseudo labels. By using the gated feature map generated by BFL, the predicted pseudo labels can enhance the attention of target features within each bounding box. Furthermore, we propose box-prompt pseudo supervision to simultaneously train BPN, PRN and MAN, which enforces the consistency between the prediction of MAN and PRN on the weakly labeled data, thus ensuring the MAN's stable optimization and maximizing the utility of the weak annotations. Comprehensive evaluations on both open-source (LIDC-IDRI) and in-house (HX-NODULE) datasets demonstrate that Box2Pseudo outperforms state-of-the-art methods and achieves comparable performance to fully-supervised approach.