自适应候选生成与多级域对抗的乳腺癌有丝分裂检测
Adaptive Candidate Generation and Multilevel Domain Adversarial for Mitosis Detection in Breast Cancer
-
摘要: 针对由于染色程序及扫描仪硬件的不统一, 用于检测的乳腺癌病理图像之间存在域偏移这一现象, 提出自适应候选生成与多级域对抗的乳腺癌有丝分裂检测方法. 首先结合图像级别与中心感知实例级别的域适应进行对抗训练, 对齐物体相关的全局和前景区域特征; 然后针对有丝分裂小目标的检测问题, 基于特征金字塔网络设计多尺度粗细回归方案, 利用动态锚点选择策略和自适应卷积挖掘不同实例大小的潜在锚点, 实现候选区域自适应; 最后, 联合分类置信度与回归精度优化, 输出高鲁棒性检测结果. 结果表明, 数据集MIDOG 2021相较于基线方法的F提高了0.029, 数据集ICPR MITOS 2014与ICPR MITOS 2012的F分别达到了0.675, 0.833, 均优于对比方法, 证明该方法对于有丝分裂检测的有效性.Abstract: To address the domain shift problem caused by inconsistent staining protocols and scanner hardware between different laboratories in mitosis detection task, this study proposes an adaptive candidate generation and multi-level domain adversarial method. An adversarial learning strategy with image-level and center-aware instance-level domain adaptation is proposed to implement alignment of global features and foreground object features. For detecting small mitotic cells, a multi-scale coarse-fine regression scheme is developed based on the feature pyramid network, which adopts a dynamic anchor selection strategy to mine potential anchors across varying instance sizes, thereby achieving adaptive region proposal generation. Finally, joint optimization of classification confidence and regression precision is conducted to produce robust detection outputs. Experimental results demonstrate performance improvements of 0.029 on MIDOG 2021 compared with baseline methods, along with achieved scores of 0.675 and 0.833 on ICPR MITOS 2014 and ICPR MITOS 2012 datasets respectively, and this proposal are superior to comparative approaches.