Segmentation of Carotid Plaques Based on Improved DeepLabV3

This paper proposes a neural network, MOGA-GCNDeepLabV3+, to improve carotid plaque segmentation in ultrasound images affected by noise, low contrast, and blurred edges. DenseNet replaces the backbone of DeepLabV3+, and MogaBlock from MOGANet enhances feature extraction. Additionally, dynamic convolutional networks (GCN) run in parallel with ASPP's dilated convolution to better model complex plaque shapes. A new loss function, TverskyLDAMLoss, combines Tversky and LDAM losses to address pixel class imbalance. Compared to DeepLabV3+, the improved model achieves 3.08% higher mIoU, 2.42% higher mDice, and 3.42% higher mAcc.