Medical image segmentation method based on multi‐scale feature and U‐net network

Abstract

AbstractIn medical image segmentation (MIS), better segmentation results can be obtained by training the deeper neural network. However, directly building too deep network will cause problems such as gradient disappearance, which will affect the segmentation effect. Therefore, a dilated inception U‐Net (DIU)‐net network is constructed by combining the multi‐scale feature fusion (MSFF) method and the concept of Inception in Google net based on U‐net, and its effectiveness is verified by experiments. The DIU‐net network's training accuracy has been improved in the lung computed tomography (CT) and fundus vascular CT image data sets. And the attenuation of the loss function is relatively stable, with the highest accuracy of 99.6%. In comparison of evaluation indicators, the values of different indicators of DIU‐net in the two data sets are higher than those of the comparison network. The DICE coefficient of DIU‐net in the lung CT image in the experiment is 0.986 on average, which is 0.2% higher than that of ResU‐net. SE value is 0.985, which is 1.9% higher than SegNet. Specificity value is slightly higher than the second segmentation effect. F1 score is 0.985, 0.6% higher than ResU‐net, area under curve value is 0.99, 0.7% higher than FCN‐8 s. In general, the DIU‐net network proposed in the study will not cause gradient disappearance and other problems in the experiment. At the same time, this method also shows high efficiency and has strong feasibility for the actual MIS.