To address the limitations of existing medical image diagnosis methods based on convolutional neural networks and Transformers, such as insufficient long-range dependency modeling and quadratic computational complexity, this paper proposes a 3D positron emission tomography (PET) image classification framework named state space hybrid convolutional model (SSHCM) . The framework is built upon a multi-scale channel space perception Mamba architecture, which integrates a linear state-space model with a multi-scale feature interaction mechanism, using stacked LMamba blocks to capture long-range dependencies in 3D voxel sequences dynamically. A layer-wise cross-scale channel attention fusion module is designed to achieve adaptive fusion of global contextual semantic. Additionally, a channel-spatial perception module is constructed by combining large kernel convolutions with an inverted bottleneck structure, enhancing spatial feature fusion and improving lesion localization accuracy. Experimental results on the Alzheimer's Disease Neuroimaging Initiative dataset with 1187 subjects show that the proposed model significantly outperforms ResNet, ViT, and Mamba variant models in terms of both accuracy and AUC. Specifically, the model achieves accuracy rates of 97.03% for AD classification and 83.33% for MCI conversion prediction tasks.

This study addressed the limitations of existing 3D point cloud completion methods, which rely on paired data and supervised learning, resulting in high data costs and limited generalization capabilities. This paper proposed a novel 3D point cloud completion method without requiring paired data for training by integrating 3D Gaussian splatting and Score Distillation Sampling (SDS) techniques. In the proposed method, incomplete point clouds were transformed into 3D Gaussian models, which were iteratively optimized using a pre-trained 2D diffusion model to predict complete point clouds and fulfill the completion task. To enhance the initial 3D Gaussian models, the authors introduced a point densification technique that increased the density of input point clouds. Additionally, a progressive camera sampling strategy was adopted during the early optimization stages to control the camera sampling range, thereby improving optimization efficiency. The experimental results demonstrate that the proposed method outperforms existing approaches on the RedWood-3DScan dataset. Ablation studies further validate the effectiveness of the optimization strategies, confirming the superiority of the proposed method in handling real-world data.