An Improved Mask R-CNN Algorithm for Localization of Pneumonia
YUAN Ya
WANG Yong
WANG Ying
Abstract：Pneumonia is one of the common lung diseases with high incidence rate.In order to solve the problem of low detec-tion accuracy of pneumonia focus location,this paper proposes an improved algorithm CS2-Mask R-CNN based on Mask R-CNN.Firstly,the hierarchical cascade feature mapping block is introduced into the backbone network,and more global feature informa-tion is obtained by increasing the receptive field.Then,in the feature pyramid network(FPN),channel and spatial dual attention mechanism(CBAM)are introduced to fuse the channel and spatial feature information,the feature dependence in the channel and spatial direction is improved,and the loss of high-level features caused by direct channel dimensionality reduction is avoided.The low level features are embedded into the high level features to enhance the transmission of feature information.In particular,for the problems of missing and wrong detection,the soft-NMS mechanism is introduced,which does not simply set the score of the box that overlaps with the highest score and is greater than the threshold to zero,but uses a penalty method to reduce its score.Finally,the algorithm is tested on the RSNA dataset,and the detection accuracy is significantly improved.
Keywords：pneumonialocalization of lesionsreceptive fieldcharacteristic pyramid networkattention mechanism
Publication Date：2025-07-20
Online Publishing Date：2025-09-25(First online date of this platform, not the publication date of the document)
Pages：7( 1817-1822,1828 )