Abstract: BACKGROUND: Rapid and accurate segmentation of brain tissue in medical images is of great significance for three-dimensional biomechanical modeling and diagnosis of craniocerebral injuries. Currently, artificial intelligence (AI)-based baseline models exhibit excellent generalization capabilities on large-scale datasets. However, due to the specificity and complexity of craniocerebral tissues, these models have certain limitations in their application to craniocerebral tissue segmentation. Additionally, the scarcity of craniocerebral tissue samples makes it difficult for baseline models to achieve precise segmentation results through fine-tuning.
OBJECTIVE: To construct a craniocerebral tissue segmentation model based on texture feature retrieval enhancement to improve segmentation accuracy under a small number of samples.
METHODS: Segment Anything in Medical Images (MedSAM) model was selected as the basic framework, and texture features were combined with deep learning to build a brain tissue segmentation model based on texture feature retrieval enhancement (DP-MedSAM). Dice Coefficient and mean intersection over union (MIoU) were selected to evaluate the efficiency of image segmentation results. In comparison with the original MedSAM model, the ablation experiment systematically evaluated the influence of key components on the model performance. The sensitivities of MedSAM, the Segment Anything Model (SAM) for medical image segmentation (SAM-Med2D) and DP-MedSAM in the mandible, left optic nerve, and left parotid gland were compared.
RESULTS AND CONCLUSION: (1) By verifying the impact of the number of point prompts on segmentation results on the HaN-Seg dataset, the experimental results indicated that the optimal Dice score was achieved with the addition of three points. (2) DP-MedSAM demonstrated performance improvements compared with MedSAM and SAM-Med2D on two datasets (HaN and Public Domain Database for Computational Anatomy). Especially on the Public Domain Database for Computational Anatomy dataset, in terms of the MIoU metric, DP-MedSAM outperformed MedSAM by 6.59% and SAM-Med2D by 37.35%; in terms of the Dice metric, DP-MedSAM outperformed MedSAM and SAM-Med2D by 4.34% and 25.32%, respectively. (3) The ablation experiment results showed that removing the texture feature extraction module in the DP-MedSAM model, relying solely on original image features, led to a significant decrease in results on the test set. Furthermore, removing the vector cache database and its retrieval enhancement function from the model, which deprived the ability of the model to perform similarity retrieval using an external knowledge base, further reduced model performance. (4) Under conditions of limited data resources, the DP-MedSAM model outperformed the other two models in all evaluation metrics. The DP-MedSAM model performed excellently when processing simple and moderately difficult samples, demonstrating a clear advantage over the other two models and indicating good generalization ability. Processing the fine structures of difficult samples placed higher demands on the model’s segmentation capabilities. Although the performance of the DP-MedSAM model declined slightly, it still outperformed the other two models. (5) This study proposes an innovative craniocerebral tissue segmentation model, DP-MedSAM, which improves the baseline model’s performance in capturing local details and global structural information in medical images by introducing target region texture feature extraction. Through vector similarity retrieval technology, DP-MedSAM can retrieve the feature vector most similar to the current target region from a pre-constructed vector database, providing more precise guiding information for the segmentation process.

Key words: craniocerebral, textural features, medical segmentation, base model, retrieval enhancement