Lesion-Aware Reconstruction with Principal Network: Enhancing Pseudo-Label Reliability in Semi-Supervised Clinical Lesion Detection

Shiwan DI, Jupeng LI, Yuxuan YANG, Qian JIN, Guorui AN, Jingwen YANG, Yue WANG, Yong GUO, Xinyue ZHANG, Ruohan MA, Gang LI
Proceedings of The 9th International Conference on Medical Imaging with Deep Learning, PMLR 315:2321-2337, 2026.

Abstract

Purpose. In lesion detection tasks, labeled medical data are often scarce, limiting the performance of fully supervised models. Teacher-student (TS) frameworks based on semi-supervised learning (SSL) have emerged as effective solutions to leverage unlabeled data. However, the inherent high-confidence bias of teacher networks frequently leads to the propagation of erroneous pseudo-labels, degrading the generalization ability of student networks. To address this critical issue, we propose a novel teacher-principal-student (TPS) framework. Methods. The core innovation lies in introducing a principal network, which integrates lesion-aware reconstruction to filter low-quality pseudo-labels generated by the teacher network. Specifically, the principal network leverages anatomical prior knowledge and reconstruction consistency constraints to assess the reliability of teacher-generated pseudo-labels, ensuring only high-fidelity pseudo-labeled data are used for training the student network. This design fundamentally mitigates the adverse effects of the teacher prediction bias and error propagation. Results. Extensive experiments on jaw lesion detection datasets demonstrate the superiority of our approach. With the same label ratio, our SSL network achieves 81.5% mAP@0.5, outperforming mainstream SSL methods by 3.0% while narrowing the performance gap with fully supervised learning to only 3.3%. Conclusion. Our proposed TPS framework outperforms state-of-the-art SSL approaches in jaw lesion detection task. It not only achieves competitive performance comparable to fully supervised models but also significantly reduces reliance on labeled clinical data, providing a reliable technical solution to promote the clinical translation of lesion detection systems.

Cite this Paper

BibTeX
@InProceedings{pmlr-v315-di26a, title = {Lesion-Aware Reconstruction with Principal Network: Enhancing Pseudo-Label Reliability in Semi-Supervised Clinical Lesion Detection}, author = {DI, Shiwan and LI, Jupeng and YANG, Yuxuan and JIN, Qian and AN, Guorui and YANG, Jingwen and WANG, Yue and GUO, Yong and ZHANG, Xinyue and MA, Ruohan and LI, Gang}, booktitle = {Proceedings of The 9th International Conference on Medical Imaging with Deep Learning}, pages = {2321--2337}, year = {2026}, editor = {Huo, Yuankai and Gao, Mingchen and Kuo, Chang-Fu and Jin, Yueming and Deng, Ruining}, volume = {315}, series = {Proceedings of Machine Learning Research}, month = {08--10 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v315/main/assets/di26a/di26a.pdf}, url = {https://proceedings.mlr.press/v315/di26a.html}, abstract = {Purpose. In lesion detection tasks, labeled medical data are often scarce, limiting the performance of fully supervised models. Teacher-student (TS) frameworks based on semi-supervised learning (SSL) have emerged as effective solutions to leverage unlabeled data. However, the inherent high-confidence bias of teacher networks frequently leads to the propagation of erroneous pseudo-labels, degrading the generalization ability of student networks. To address this critical issue, we propose a novel teacher-principal-student (TPS) framework. Methods. The core innovation lies in introducing a principal network, which integrates lesion-aware reconstruction to filter low-quality pseudo-labels generated by the teacher network. Specifically, the principal network leverages anatomical prior knowledge and reconstruction consistency constraints to assess the reliability of teacher-generated pseudo-labels, ensuring only high-fidelity pseudo-labeled data are used for training the student network. This design fundamentally mitigates the adverse effects of the teacher prediction bias and error propagation. Results. Extensive experiments on jaw lesion detection datasets demonstrate the superiority of our approach. With the same label ratio, our SSL network achieves 81.5% mAP@0.5, outperforming mainstream SSL methods by 3.0% while narrowing the performance gap with fully supervised learning to only 3.3%. Conclusion. Our proposed TPS framework outperforms state-of-the-art SSL approaches in jaw lesion detection task. It not only achieves competitive performance comparable to fully supervised models but also significantly reduces reliance on labeled clinical data, providing a reliable technical solution to promote the clinical translation of lesion detection systems.} }
Endnote
%0 Conference Paper %T Lesion-Aware Reconstruction with Principal Network: Enhancing Pseudo-Label Reliability in Semi-Supervised Clinical Lesion Detection %A Shiwan DI %A Jupeng LI %A Yuxuan YANG %A Qian JIN %A Guorui AN %A Jingwen YANG %A Yue WANG %A Yong GUO %A Xinyue ZHANG %A Ruohan MA %A Gang LI %B Proceedings of The 9th International Conference on Medical Imaging with Deep Learning %C Proceedings of Machine Learning Research %D 2026 %E Yuankai Huo %E Mingchen Gao %E Chang-Fu Kuo %E Yueming Jin %E Ruining Deng %F pmlr-v315-di26a %I PMLR %P 2321--2337 %U https://proceedings.mlr.press/v315/di26a.html %V 315 %X Purpose. In lesion detection tasks, labeled medical data are often scarce, limiting the performance of fully supervised models. Teacher-student (TS) frameworks based on semi-supervised learning (SSL) have emerged as effective solutions to leverage unlabeled data. However, the inherent high-confidence bias of teacher networks frequently leads to the propagation of erroneous pseudo-labels, degrading the generalization ability of student networks. To address this critical issue, we propose a novel teacher-principal-student (TPS) framework. Methods. The core innovation lies in introducing a principal network, which integrates lesion-aware reconstruction to filter low-quality pseudo-labels generated by the teacher network. Specifically, the principal network leverages anatomical prior knowledge and reconstruction consistency constraints to assess the reliability of teacher-generated pseudo-labels, ensuring only high-fidelity pseudo-labeled data are used for training the student network. This design fundamentally mitigates the adverse effects of the teacher prediction bias and error propagation. Results. Extensive experiments on jaw lesion detection datasets demonstrate the superiority of our approach. With the same label ratio, our SSL network achieves 81.5% mAP@0.5, outperforming mainstream SSL methods by 3.0% while narrowing the performance gap with fully supervised learning to only 3.3%. Conclusion. Our proposed TPS framework outperforms state-of-the-art SSL approaches in jaw lesion detection task. It not only achieves competitive performance comparable to fully supervised models but also significantly reduces reliance on labeled clinical data, providing a reliable technical solution to promote the clinical translation of lesion detection systems.
APA
DI, S., LI, J., YANG, Y., JIN, Q., AN, G., YANG, J., WANG, Y., GUO, Y., ZHANG, X., MA, R. & LI, G.. (2026). Lesion-Aware Reconstruction with Principal Network: Enhancing Pseudo-Label Reliability in Semi-Supervised Clinical Lesion Detection. Proceedings of The 9th International Conference on Medical Imaging with Deep Learning, in Proceedings of Machine Learning Research 315:2321-2337 Available from https://proceedings.mlr.press/v315/di26a.html.

Related Material