Segmenting What Matters: A Dual Stage Active Learning Framework for Weakly Supervised Breast Ultrasound Segmentation

Accurate tumor segmentation in breast ultrasound is essential for effective early detection of breast cancer. While fully supervised segmentation methods require extensive pixellevel annotations, we aim to address this limitation by proposing a weakly supervised segmentation framework using image-level labels. Our pipeline builds upon the general framework of class activation map (CAM)-based weakly supervised image segmentation, which typically involves training a classification model, generating CAMs, creating pseudo pixel-level labels from the CAMs, and training a segmentation model on these pseudo labels. However, this conventional approach faces several limitations: (1) CAM quality is poor for certain samples; (2) not all images yield reliable CAMs; and (3) segmentation performance suffers due to low-quality pseudo labels. To address these issues, our contributions are organized into three key components: (1) We propose an active learning approach that selects informative samples based on predictive entropy to boost CAM quality. (2) Instead of using CAMs from all images for pseudo-label generation, we apply an HSV-based CAM filtering technique to identify high-quality, tumor-relevant CAMs. (3) We train the segmentation model using the Mean Teacher framework, leveraging both the selected pseudo-labeled images and the remaining unlabeled images. This framework is further enhanced by a second round of active learning, which selects the most uncertain samples for refinement. Experimental results on two public breast ultrasound datasets show that our method achieves superior accuracy and outperforms most existing state-of-the-art methods, while significantly reducing reliance on full supervision. The code is available at https://github.com/Steven-ZN/DSAL.