Shape prior-constrained deep learning network for medical image segmentation.

We propose a shape prior representation-constrained multi-scale features fusion segmentation network for medical image segmentation, including training and testing stages. The novelty of our training framework lies in two modules comprised of the shape prior constraint and the multi-scale features fusion. The shape prior learning model is embedded into a segmentation neural network to solve the problems of low contrast and neighboring organs with intensities similar to the target organ. The latter can provide both local and global contexts to address the issues of large variations in patient postures as well as organ's shape. In the testing stage, we propose a circular collaboration framework strategy which combines a shape generator auto-encoder network model with a segmentation network model, allowing the two models to collaborate with each other, resulting in a cooperative effect that leads to accurate segmentations. Our proposed method is evaluated and demonstrated on the ACDC MICCAI'17 Challenge Dataset, CT scans datasets, namely, in COVID-19 CT lung, and LiTS2017 liver from three different datasets, and its results are compared with the recent state of the art in these areas. Our method ranked 1st on the ACDC Dataset in terms of Dice score and achieved very competitive performance on COVID-19 CT lung and LiTS2017 liver segmentation.

Light3DHS: A lightweight 3D hippocampus segmentation method using multiscale convolution attention and vision transformer.

The morphological analysis and volume measurement of the hippocampus are crucial to the study of many brain diseases. Therefore, an accurate hippocampal segmentation method is beneficial for the development of clinical research in brain diseases. U-Net and its variants have become prevalent in hippocampus segmentation of Magnetic Resonance Imaging (MRI) due to their effectiveness, and the architecture based on Transformer has also received some attention. However, some existing methods focus too much on the shape and volume of the hippocampus rather than its spatial information, and the extracted information is independent of each other, ignoring the correlation between local and global features. In addition, many methods cannot be effectively applied to practical medical image segmentation due to many parameters and high computational complexity. To this end, we combined the advantages of CNNs and ViTs (Vision Transformer) and proposed a simple and lightweight model: Light3DHS for the segmentation of the 3D hippocampus. In order to obtain richer local contextual features, the encoder first utilizes a multi-scale convolutional attention module (MCA) to learn the spatial information of the hippocampus. Considering the importance of local features and global semantics for 3D segmentation, we used a lightweight ViT to learn high-level features of scale invariance and further fuse local-to-global representation. To evaluate the effectiveness of encoder feature representation, we designed three decoders of different complexity to generate segmentation maps. Experiments on three common hippocampal datasets demonstrate that the network achieves more accurate hippocampus segmentation with fewer parameters. Light3DHS performs better than other state-of-the-art algorithms.

Pedestrian re-identification based on attention mechanism and Multi-scale feature fusion.

Existing pedestrian re-identification models generally have low pedestrian retrieval accuracy when encountering factors such as changes in pedestrian posture and occlusion because the network cannot fully express pedestrian feature information. Therefore, this paper proposes a method to address this problem by combining the attention mechanism with multi-scale feature fusion, and combining the proposed cross-attention module with the ResNet50 backbone network. In this way, the ability of the network to extract strong salient features is significantly improved; at the same time, using the multi-scale feature fusion module to extract multi-scale features from different depths of the network, achieving the complementary advantages between features through feature addition, feature concatenation and feature weight selection. In addition, a feature enhancement method and an efficient pedestrian retrieval strategy are proposed to jointly promote the accuracy of pedestrian retrieval from both the training and testing levels. When tested on the occluded pedestrian recognition datasets Partial-REID and Partial-iLIDS, the accuracy of this method reached 70.1% and 65.6% on the Rank-1 indicator respectively, and 82.2% and 80.5% on the Rank-3 indicator respectively. At the same time, it also achieved high recognition accuracy when tested on the Market1501 dataset and DukeMTMC-reid dataset, reaching 95.9% and 89.9% on the Rank-1 indicator respectively, 89.1% and 80.3% on the mAP indicator respectively, and 67% and 46.2% on the mINP indicator respectively. It can be seen that this method has achieved good results in solving the above problems.