The high annotation costs in medical detection tasks make semi-supervised object detection (SSOD) methods particularly promising. However, current research has paid little attention to medical image SSOD, and conventional SSOD approaches struggle with the unique challenges of medical imaging—scarce annotations and difficult-to-learn features.
Prof. Lei ZHANG from Sichuan University, China, and her students Jiaqiang CHEN et al. were dedicated to improving the efficient utilization of medical image annotations, and published their new research on 15 June 2026 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.
When applying state-of-the-art SSOD methods to medical images, the research team observed significant performance degradation compared to expected results. Through visual analysis, they identified the core issue lying in these methods' pseudo-labeling mechanisms. Comparative studies with natural image experiments revealed that conventional pseudo-labels in medical imaging scenarios weaken consistency constraints and bias box regression.
The research team revealed that mainstream detectors (both two-stage and single-stage) universally generate predictions based on priors boxes (anchors or proposals) during systematic analysis of their computational workflows. Post-processing (NMS and confidence thresholding) are required to obtain final detection results. This inspired them to propose Priors-Level Pseudo-Labels (PLPLs). PLPLs refer to the raw model outputs before post-processing. Since they maintain one-to-one correspondence with priors boxes, the outputs of teacher and student models can be aligned through this prior-based mapping. By applying a transformation from the teacher's view to the student's view, the teacher's outputs can directly guide student model training. Compared to common pseudo-labels, PLPLs strengthen consistency constraints by directly aligning teacher and student outputs. Additionally, the negative impact of inaccurate boxes on box regression is reduced, as each box only affects the refinement of a single prior box. Combined with the reweighting mechanism in the PriorsMatch framework, this approach enhances pseudo-label utilization efficiency and further improves detection performance.
The team then turned their attention to the annotation process of medical images, aiming to select information-rich samples during the initial dataset construction phase. This approach is designed to enhance annotation efficiency and reduce clinicians' workload. Their research demonstrates remarkable continuity in addressing core challenges throughout these investigations.

DOI:10.1007/s11704-025-41199-2