从局部到区域分层的乳腺病理图像有丝分裂检测

Mitotic Detection in Breast Histopathology Images Using Local and Regional Hierarchical Information

统计乳腺癌组织学图像中有丝分裂细胞的数量是乳腺癌分级和预后的重要诊断依据。目前有丝分裂细胞计数主要由病理学家手工进行,是一项费时费力的任务。为解决这一具有挑战性的有丝分裂细胞检测问题,本研究提出了一种从局部到区域分层的乳腺癌病理学图像有丝分裂检测方法。框架整体由两阶段构成,第一阶段为细胞定位网络,从整切片图像中筛查、定位候选的有丝分裂细胞图像块,同时引入深监督机制与解耦的检测头来提升性能;第二阶段为有丝细胞验证网络,负责对大量的候选细胞图像块进一步细化分类,使用基于图注意机制的上下文融合网络,通过整合大范围的区域特征来调节局部中心块的原有响应,从而得到更准确分类结果。在ICPR MITOSIS 2014、ICPR MITOSIS 2012和TUPAC16数据集上分别使用960、35和649个高倍视野图像(HPF)作为训练集,240、15和7个HPF作为测试集,分别取得0.676、0.809和0.797的F-Score,其中召回率均取得了最优结果,分别为0.878、0858、0.875。所提出的有丝分裂自动检测方法能够高效的检测病理切片中的癌细胞,具有良好的临床应用价值。

Statistically counting the number of mitotic cells in histological images of breast tumor tissue is an important diagnostic basis for the grading and prognosis of breast cancer.Currently,the counting tasks are performed manually by pathologists,which is a time-consuming and laborious task.To address this challenge,this paper proposed a method for mitotic detection in breast cancer pathology images from local to regional stratification.The framework as a whole consisted of two stages.The first stage was a cell localization network,which was responsible for screening and locating candidate mitotic cell blocks from whole section images while introducing a deep supervision mechanism with decoupled detection heads to enhance performance.The second stage was the mitotic cell validation network,which was responsible for further refining the classification of a large number of candidate cell image blocks by using a contextual fusion network based on a graph-attention mechanism to modulate the original response of local central blocks by integrating a large range of regional features to obtain more accurate classification results.We achieved F-Scores of 0.676,0.809,and 0.797 on the ICPR MITOSIS 2014,ICPR MITOSIS 2012,and TUPAC16 datasets,respectively,using 960,35,and 649 HPF as training sets,and 240,15,and 7 HPF as test sets,respectively,where the recall rates all achieved optimal results of 0.878,0.858 and 0.875,respectively.The results indicated that the proposed automatic detection method efficiently detected the cancer cells in the pathological sections,showing significant clinical application value.