Cross-layer Channel Pruning Method Based on Variable Sequences

Radiotherapy is an important treatment modality for liver cancer. Deep learning-based image semantic segmentation technology can assist physicians in demarcating radiation target areas and enhancing the accuracy of radiotherapy. However, existing medical image semantic segmentation models are relatively intricate and possess a substantial number of parameters, rendering them challenging to deploy on devices with constrained resources. An analysis of the significance of the parameters of the vision transformer model reveals that the crucial parameters of the different layers of the model exhibit a distinctive distribution pattern. Based on this finding, this study proposes a cross-layer channel pruning method based on variable sequences. According to the distribution pattern of the significant parameters, the significance weights of the Multihead Self-Attention (MSA) and Feed-Forward Network (FFN) layers are measured and these values are adjusted to form a hierarchical sequence of significance weight values. Subsequently, the corresponding pruning rate is set for the sequence to form a variable pruning rate sequence that varies with the depth of the network, thereby achieving fine pruning of the MSA and FFN layers. This new method introduces a cyclic pruning strategy that iteratively updates the variable pruning rate sequence during each round of model pruning to reduce the redundant structures in the MSA and FFN layers adequately. The model is trained and tested using the public liver segmentation dataset, 3D-IRCADb-01. After pruning the vision transformer, the accuracy of the image segmentation does not decline and the Floating-Point Operations (FLOPs) and number of parameters are reduced by 60.26% and 66.07%, respectively. Experimental results indicate that the new method attains a higher pruning rate while guaranteeing segmentation accuracy and is more advantageous than the fixed pruning rate method.