The inherent “black-box” nature of neural network models leads to opaque internal logic and behavioral decision， restricting their application in critical areas. To address this issue， based on the Layer-wise Relevance Propagation （LRP） method， the interpretability of image classification models such as VGG16 and ResNet50 was researched， and an interpretation method for image classification models named SSGLRP （Selective Softmax Gradient for Layer-wise Relevance Propagation） was proposed. By introducing activation values for the positive gradients of output neurons and modifying the initial relevance values for non-target classes， the problem of LRP heatmaps containing noise and lacking class discrimination was effectively solved. The SSGLRP method was quantitatively evaluated using maximum patch masking and fixed-point game experiments， with classic LRP， Selective Layer-wise Relevance Propagation （SLRP）， and Softmax Gradient Layer-wise Relevance Propagation （SGLRP） as baseline methods. The results of the maximum patch masking experiments show that on VGG16， the prediction change obtained by SSGLRP is on average 91.8%， 60.3%， and 34.2% higher than those obtained by LRP， SLRP， and SGLRP， respectively； on ResNet50， the improvements are 61.9%， 64.9%， and 33.2%， respectively. The SSGLRP method has higher class discrimination ability and less noise with superior performance in interpreting VGG16 and ResNet50 models.

To solve target tracking with target changes under complex background, an adaptive target tracking method that combined graph-based semi-supervised learning method with the particle filter was proposed. It used LLGC (Learning with Local and Global Consistency) algorithm to establish the cost function, and took current status of the candidate as unlabeled samples, then established diagram using all samples as vertex, taking the optimal solution of the cost function as current status, obtaining the target position in current frame. Besides, it used the tracking result to update the labeled samples in real time, so that the algorithm could adapt to the target deformation, partial occlusion and illumination changes. Analysis and experiment show that the proposed method can handle complicated situations like occlusion or similar background interference very well, and achieves target tracking robustly.

To acquire an accurate patient positioning in image-guided radiotherapy, an improved Demons deformable registration method was developed. The FDK algorithm was adopted to reconstruct Cone Beam CT (CBCT) and the reconstruction result was visualized by a volume rendering method with Visualization ToolKit (VTK). Based on the Insight segmentation and registration ToolKit (ITK), the Demons algorithm was completed incorporating the gradient information of fixed image and floating image by the concept of symmetric gradient, and a new formula of Demons force was demonstrated. Registrion experiments were carried out using medical images both from single modality and multi-modality. The results show that the improved Demons algorithm achieves a faster convergence speed and a higher precision compared with the original demons algorithm, which indicates that the Demons algorithm based on symmetric gradient is more suitable for the registration of CBCT reconstruction image and CT plan image in image-guided radiotherapy.