A Lightweight Complex Image Recognition Method Based on MCRAIP-Net
Haotian Liu, Longtian Fu, Qi Zhang, Lizhi Lin, Ruel Reyes · IEEE Access · 2024
To improve the accuracy and time efficiency of medical brain image segmentation, we propose a deformable registration network EDUNet in the registration stage of multi-atlas segmentation (MAS) of brain images. We perform data preprocessing on the floating and fixed images to minimize external influences. In the registration stage, we use ANTs instead of traditional “coarse” registration, and employ a convolutional neural network (CNN) to improve “fine” registration, estimate the deformation field, and introduce an attention mechanism and a dilated convolution module. We propose a deep learning based multimodal cross-reconstruction inverted pyramid network MCRAIP-Net, which uses multimodal magnetic resonance images as input, extracts features of each modality through three independent encoder structures, and fuses the extracted features at the same resolution level. We use a dual-channel cross-reconstruction attention module to refine and fuse multimodal features. On this basis, we use an inverted pyramid decoder to integrate the features of different resolutions at each stage of the decoder, and complete the task of brain tissue segmentation. We compare the segmentation results of two datasets, and show that the proposed algorithm has simpler interaction and faster speed than the confidence connected algorithm, and also greatly improves the segmentation performance. On the Synapse dataset, we use 18 samples as training sets and 12 samples as testing sets; 5 On the ACDC dataset, we use 70 samples as training set, 10 samples as validation set, and 20 samples as testing set. The experimental results show that compared with the traditional U-Net network and other existing models, our model has better accuracy and feasibility in rib fracture segmentation.