Dendritic U-Net for robust vascular wall segmentation in ultrasound imaging.

PURPOSE: We proposed a method that combines the deep learning model U-Net with a dendritic neuron model (DNM) and demonstrated its effectiveness for more robust extraction of vascular wall boundaries through analysis using ultrasound simulations and in vivo carotid artery data. METHODS: Dendritic U-Net (DU-Net) is a model that replaces the output layer of U-Net with a DNM. Its accuracy was compared to U-Net using cross-validation on a dataset generated by ultrasound simulation. The applicability of the model trained on the simulation dataset to clinical data was also verified using an in vivo dataset of a healthy human carotid artery. The noise robustness of the results was evaluated using the mean absolute difference (MAD) within the scanline and frame of the extracted vessel wall boundaries from the segmentation map. RESULTS: Cross-validation using the simulation data showed that DU-Net achieved higher accuracy than U-Net in terms of precision (93.64%), Dice (96.38%), and Intersection over Union (93.01%). Furthermore, evaluation of the MAD of the extracted vessel wall boundaries from the in vivo data demonstrated that DU-Net produced smoother boundaries than U-Net. CONCLUSION: Incorporating DNM into U-Net suppresses the impact of noise on vessel wall boundary extraction, demonstrating the effectiveness of the proposed DU-Net method for vessel wall boundary extraction.