机电工程技术2025，Vol.54Issue(24)：26-31,158,7.DOI:10.3969/j.issn.1009-9492.2025.24.005
多分支混合监督前列腺癌格里森分级方法
Multi-branch Mixed Supervision Method for Gleason Grading of Prostate Cancer
摘要
Abstract
Current Gleason grading methods based on multi-instance learning predominantly rely on weak supervision with slide-level labels,which often results in substantial performance degradation due to insufficient supervision.Although some hybrid supervision strategies integrate both slice-level and pixel-level labels,their accuracy remains suboptimal as the specific discriminative mechanisms of Gleason grading are overlooked.To address these limitations,a multi-branch hybrid supervised Gleason grading approach is proposed.Limited pixel-level labels are converted into instance-level labels,and a hybrid supervised model incorporating both instance-level and slide-level labels is employed to improve classification performance.A multi-branch structure informed by clinical practice is constructed to handle different Gleason grades specifically,reducing computational resource consumption while more effectively preserving the integrity of cancer lesion information,thereby enhancing the accuracy for malignant grading.The proposed method achieves QWK and ACC scores of 0.867 and 0.817 on the PANDA dataset and 0.942 and 0.859 on the PUMCH dataset,outperforming advanced multi-instance learning methods such as TransMIL and MSMIL.关键词
多实例学习/格里森分级/弱监督/混合监督Key words
multi-instance learning/Gleason grading/weak supervision/hybrid supervision分类
信息技术与安全科学引用本文复制引用
曾军英,尤吴杭,邓森耀,麦智鹏,严维刚,肖雨,秦传波,贾旭东..多分支混合监督前列腺癌格里森分级方法[J].机电工程技术,2025,54(24):26-31,158,7.基金项目
中央高水平医院临床科研专项(2022-PUMCH-B-009) （2022-PUMCH-B-009）
2022年研究生教育创新计划项目(YJS-SFJD-22-01) （YJS-SFJD-22-01）
2024年广东省普通高校重点领域专项(2024ZDZX1008) （2024ZDZX1008）
