基于高分特征保留网络的中医齿痕舌图像分类方法

Image Classification Method for Tooth-Marked Tongue in Traditional Chinese Medicine Based on High-Resolution Feature Preservation Network

摘要: 齿痕舌是反映脾虚、水湿内停等病理状态的重要中医客观体征。现有深度学习模型在提取此类局限于舌体边缘的微小形态特征时，常因网络过深导致“空间平滑效应”，且全局平均池化(GAP)易稀释微弱的异常信号，引发临床漏诊。为此，本文提出一种基于高分特征保留网络(HRFP-Net)的齿痕舌图像分类方法。该方法引入早期退出(Early-Exit)机制，在轻量化骨干网络(ConvNeXt-Tiny)浅层进行截断，以保留高分辨率的物理边缘形态；同时采用全局最大池化(GMP)构建峰值信号检测器，精准锁定并提取局部异常激活区域。在公开的Tooth-Marked数据集上的实验表明，HRFP-Net有效克服了边缘特征被掩盖的问题，准确率达94.42%，精确率高达97.06%，F₁分数和AUC分别为93.40%和97.73%，性能显著优于主流的CNN与Transformer模型。本文方法在降低网络冗余的同时实现了微小齿痕体征的高敏锐度捕捉，为中医客观化辅助诊断提供了高效、鲁棒的新思路。

Abstract: The tooth-marked tongue is a crucial objective sign in Traditional Chinese Medicine (TCM) that reflects pathological states such as spleen deficiency and the internal retention of dampness and fluid. When extracting such subtle morphological features confined to the edge of the tongue, existing deep learning models often suffer from a “spatial smoothing effect” due to excessive network depth. Furthermore, Global Average Pooling (GAP) tends to dilute weak abnormal signals, leading to missed clinical diagnoses. To address these issues, this paper proposes an image classification method for tooth-marked tongues based on a High-Resolution Feature Preservation Network (HRFP-Net). This method introduces an Early-Exit mechanism that truncates the lightweight backbone network (ConvNeXt-Tiny) at a shallow layer to preserve high-resolution physical edge morphologies. Simultaneously, Global Max Pooling (GMP) is employed to construct a peak signal detector, which accurately locks onto and extracts locally abnormal activated regions. Experiments on the public Tooth-Marked dataset demonstrate that HRFP-Net effectively overcomes the problem of edge feature obfuscation. It achieves an accuracy of 94.42% and a remarkably high precision of 97.06%, with an F1-score and AUC of 93.40% and 97.73%, respectively. Its overall performance is significantly superior to mainstream CNN and Transformer models. While reducing network redundancy, the proposed method achieves high-sensitivity capture of subtle tooth-marked signs, providing an efficient and robust new approach for objective computer-aided diagnosis in TCM.