摘要
针对目前慢性肾病诊断耗时耗力,人工诊断困难以及识别、分割准确率较低等情况,提出一种能够结合肾小球滤过率指标以及图像纹理特征进行综合分析的实例分割模型MBF RCNN,并提高了在慢性肾病分期上检测分割的准确率。针对基线网络识别精度较低的情况,提出特征融合模块,自适应地融合多尺度特征层,并降低不同层次特征图的不一致性,提升检测的准确率。针对分割边界模糊问题,融入自注意力机制,以充分学习图像的空间特征;加入边界细化模块,进一步优化分割的边界。模型使用新的组合损失函数提高模型的准确度。建立慢性肾病数据集,并在此数据集上与其他先进的实例分割模型进行了比较,在mAP上比基线网络Mask RCNN提升约3百分点,在分割精度上比基线提升了3.46百分点。
In view of the current time-consuming and labor-intensive diagnosis of chronic kidney disease,the difficulty of manual diagnosis,and the low accuracy of recognition and segmentation,an instance segmentation model MBF RCNN,which can combine the glomerular filtration rate index and the comprehensive analysis of image texture features,is proposed to improve the accuracy of detection and segmentation in chronic kidney disease staging.Aimed at the low recognition accuracy of the baseline network,a feature fusion module was proposed,which adaptively fused multi-scale feature layers,reduced the inconsistency of feature maps at different levels,and improved the detection accuracy.For the blurred segmentation boundary,a self-attention mechanism was incorporated to fully learn the spatial features of the image.A boundary refinement module was added to further optimize the segmentation boundary.The model used a new combined loss function to improve the accuracy of the model.A chronic kidney disease dataset was established,and we compared this model with other state-of-the-art instance segmentation models on this dataset.It is about 3 percentage points higher than the baseline network Mask RCNN on mAP and 3.46 percentage points higher than the baseline in segmentation accuracy.
作者
王茹
赵希梅
Wang Ru;Zhao Ximei(College of Computer Science and Technology,Qingdao University,Qingdao 266071,Shandong,China;Shandong Province Key Laboratory of Digital Medicine and Computer Aided Surgery,Qingdao 266071,Shandong,China)
出处
《计算机应用与软件》
北大核心
2025年第8期242-252,272,共12页
Computer Applications and Software
基金
国家自然科学基金青年科学基金项目(62006134)。
关键词
慢性肾病
FFM
边界细化模块
自注意力机制
实例分割
Chronic kidney disease
FFM
Boundary refinement module
Self-attention mechanism
Instance segmentation
