改进Hybrid-Task-Cascade的染色体分割研究  

Research on Improving Chromosome Segmentation in Hybrid Task Cascade

作　　者：彭文[1] 许树颖 PENG Wen;XU Shu-ying(Department of Control and Computer Engineering,North China Electric Power University,Beijing 102206,China)

机构地区：[1]华北电力大学控制与计算机工程学院,北京102206

出　　处：《计算机仿真》2024年第6期267-273,共7页Computer Simulation

摘　　要：针对人工分割染色体图像中的实例存在耗时长,精度不佳等问题,提出一种基于改进Hybrid Task Cascade模型的染色体实例分割方法。首先,提出基于实例操作的染色体增强策略,以扩充量少且信息不丰富的染色体数据集。然后使用PAFPN代替Hybrid Task Cascade中的特征金字塔模块,保留浅层特征信息,提高定位和分割染色体实例的准确度。针对重叠的染色体簇引入Soft-NMS方法改进候选框的筛选,保留更多的染色体包围框。最后,将测试集的结果与其它模型进行对比,采用平均准确率(mean average precision, mAP)、AP50和AP75评估模型定位和分割性能,通过对自采集的染色体图像进行评估验证,其包围框定位平均准确率和分割平均准确率分别达到了80.53%和77.55%。实验表明,上述方法在染色体图像数据集上具有较好的分割效果。Aiming at the problems of time-consuming and poor accuracy in manually segmenting instances in chromosome images,a chromosome instance segmentation method based on improved Hybrid Task Cascade model was proposed.Firstly,a chromosome enhancement strategy based on case operation was proposed to expand the chromo⁃some data set with small amount and insufficient information.Then PAFPN was used to replace the feature pyramid module in Hybrid Task Cascade to retain the shallow feature information and improve the accuracy of locating and segmenting chromosome instances.For overlapping chromosome clusters,Soft-NMS method was introduced to improve the screening of candidate frames and retain more chromosome bounding boxes.Finally,the results of the test set were compared with other models.The average accuracy(mAP),AP50 and AP75 were used to evaluate the performance of the model.Through the evaluation and verification of the self-collected mitotic metaphase chromosome microscopic images,the average accuracy of bounding box positioning and segmentation are 80.71%and 76.89%respectively.Experiments show that this method has good segmentation effect on chromosome image dataset.

关 键 词：实例分割 目标检测 数据增强 染色体分割 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]