Convolution Neural Network for Classification L3 Slice and Segmentation of Psoas in Patients with Aortic Dissection

Aortic dissection (AD) is one of the life-threatening cardiovascular emergencies, and early prediction facilitates patient management and treatment. Current research is focused on segmenting and quantifying the true and false lumen of the aorta, but aortic dissection is a rare condition with little published data. Recently the psoas muscle has been recognized as a predictor of prognosis in cardiovascular surgery as well, and there are fewer studies in this area. In this study, we used our own dataset to compare the effect of pre-training VGG16, InceptionV3 and ResNet50 networks for binary classification of third lumbar vertebrae (L3) slices as well as optimizing the psoas muscle segmentation model with residual connectivity or convolutional block attention modules (CBAM) to the U-Net network. The results showed that VGG16 classifies the best. The Dice coefficients of left and right psoas in the best segmentation model was 0.940 ± 0.052 and 0.941 ± 0.036, respectively.