基于深度学习的舌象特征研究
The Research of Tongue Features Base on Deep Learning
目的/意义 本研究将深度学习技术应用于舌象分析中，以达到舌象分析自动化的目的，并为中医舌象标准化、客观性提供参考依据，进一步推进中医诊疗技术现代化进程。方法/过程 首先构建了一个新的以区域关联性为基础引入标签松弛技术的语义分割损失函数，显式约束舌象分割模型学习局部区域各像素关联性的同时，对错误标签一定的冗余能力，其次从舌象特征中内含的是否颜色相关的潜在先验出发，在模型构建阶段将舌象特征仅解耦为两类下游多标签分类任务，在加速模型拟合的同时有效降低了模型的复杂度。结果/结论 最后，在自建数据集上，验证了本文所提算法的有效性，舌象分割MIoU指标为96.57%，舌象分析宏F1值、平均准确率分别为88.58%、82.59%。
中医学医学研究方法计算技术、计算机技术
迁移学习舌象特征深度学习舌体分割中医舌诊
何佳俊,崔涛,李睿,垢德双,赵亮,何华.基于深度学习的舌象特征研究[EB/OL].(2024-02-18)[2025-10-05].https://www.biomedrxiv.org.cn/article/doi/bmr.202404.00020.点此复制
Purpose/Significance The present research utilizes deep learning methodologies within the context of tongue manifestation analysis, with the objective of automating the process and contributing to the establishment of standardized and objective benchmarks for tongue diagnosis in Traditional Chinese Medicine (TCM). This work endeavors to significantly propel the progression towards the modernization of TCM diagnostic procedures. Method/Process Firstly, a new semantic segmentation loss function has been devised that fundamentally integrates regional association and label relaxation techniques. This advanced function imposes a constraint on the segmentation model such that it learns to account for the inter-pixel relationships within specific regions of the tongue images, while simultaneously endowing the model with a degree of resilience against inaccurately annotated labels. Secondly, leveraging the latent color-related priors inherently present within the characteristics of tongue images, a strategic decision was made to decompose these features exclusively into two separate streams for subsequent multi-label classification tasks during the models architectural design stage. This strategic maneuver serves to expedite the models convergence process while significantly diminishing its overall complexity. Result/Conclusion Lastly, the effectiveness of the algorithm proposed in this paper was validated through experimentation on our custom-built dataset. The results showed an impressive 96.57% Mean Intersection over Union (MIoU) score for tongue image segmentation, along with macro F1-score and average accuracy values of 88.58% and 82.59%, respectively.
Transfer LearningTongue Image FeaturesDeep LearningTongue SegmentationTongue Diagnosis of Traditional Chinese Medicine
展开英文信息
评论