基于舌象的桥本甲状腺炎智能鉴别研究

Research on Intelligent Identification of Hashimoto’s Thyroiditis Based on Tongue Images

摘要: 桥本甲状腺炎是免疫相关的甲状腺炎性疾病，发病率逐年升高。甲状腺功能减退(简称“甲减”)是桥本甲状腺炎最常见的并发症，舌象作为中医诊断的重要依据，能够反映机体的整体生理与病理状态，在疾病无创筛查中具有潜在应用价值。针对传统舌诊方法依赖人工经验，主观性较强且缺乏标准化量化分析手段等问题，本文提出了一种基于深度学习的甲减舌象二分类智能诊断方法。首先，对舌体区域裁剪并做颜色归一化及多尺度数据增强等预处理策略，以提升模型对关键舌象特征的表征能力。在模型设计方面，选取EfficientNet-B0和ResNet两种代表性深度学习架构，在统一实验设置下进行系统对比，并从分类精度、推理效率及泛化能力等多个维度进行综合评估。实验结果表明，所提出方法在不同模型上均取得良好性能，其中ResNet在综合性能上表现最优，在测试集上取得较高分类准确率，并在精度与计算效率之间实现良好平衡。同时，在小样本训练条件下，模型仍能够保持稳定的分类表现，体现出良好的泛化能力。本研究表明，基于深度学习的舌象分析方法能够有效提升甲减筛查的客观性与准确性。所提出系统具备无创、低成本及可扩展等优势，尤其适用于基层医疗场景，为甲状腺疾病的早期筛查与智能化辅助诊断提供了一种具有重要临床应用潜力的技术路径。

Abstract: Hashimoto’s thyroiditis is an immune-related thyroid inflammatory disease, and its incidence has been increasing year by year. Hypothyroidism (commonly referred to as “hypothyroidism”) is the most common complication of Hashimoto’s thyroiditis. Tongue diagnosis, as an important basis for traditional Chinese medicine (TCM) diagnosis, can reflect the body’s overall physiological and pathological status. It holds potential value for non-invasive screening in diseases. To address the problems with traditional tongue diagnosis methods, which rely on human experience, are subjective, and lack standardized quantitative analysis tools, this paper proposes a deep learning-based intelligent diagnosis method for hypothyroidism tongue images with binary classification. First, the tongue region is cropped, followed by color normalization and multi-scale data augmentation preprocessing strategies to enhance the model’s ability to represent key tongue image features. For model design, two representative deep learning architectures, EfficientNet-B0 and ResNet, are selected and systematically compared under a unified experimental setup. The comparison includes evaluation in multiple dimensions, such as classification accuracy, inference efficiency, and generalization ability. Experimental results show that the proposed method performs well across different models, with ResNet achieving the best overall performance, attaining high classification accuracy on the test set, and achieving a good balance between precision and computational efficiency. Additionally, under small sample training conditions, the model still maintains stable classification performance, demonstrating good generalization ability. This study demonstrates that deep learning-based tongue image analysis can significantly improve the objectivity and accuracy of hypothyroidism screening. The proposed system offers advantages such as non-invasiveness, low cost, and scalability. It is especially suitable for primary healthcare settings, providing a technology path with significant clinical application potential for early screening and intelligent auxiliary diagnosis of thyroid diseases.