ACCURATE CT PNEUMOTHORAX SEGMENTATION AND VOLUME QUANTIFICATION BASED ON ATTENTION MIX-LOSS

The timely diagnosis of CT pneumothorax is particularly important. In this paper, an automatic segmentation and quantification method for CT pneumothorax is proposed. The threshold method was used to segment lung field to remove the influence of complex environment in lung field on CT pneumothorax segmentation. The pneumothorax was segmented based on Ma_Unet (Mixloss attention U_Net), and the shape, size and location information of the target area were automatically learned to alleviate the problem of data imbalance and optimize network training. The pneumothorax volume and lung compression ratio algorithm were proposed to achieve pixel level accurate CT pneumothorax volume quantification. The experimental results on the created Seg-CT-Pne test set show that the accuracy of proposed method is 99.96%, which is superior to the existing threshold method, U_Net and nnU-Net, and realizes the CT pneumothorax volume quantitative and lung compression ratio calculation. The whole automatic segmentation and quantization process only takes 18.87 s on average, and the mean difference in CT pneumothorax volume is only 4.47%, which meet the clinical needs.