基于改进ResNet34的皮肤病分类研究

Research on Skin Disease Classification Based on Improved ResNet34

摘要: 针对皮肤镜图像分类中深层卷积神经网络特征提取不充分及梯度消失的问题，本文提出了一种基于改进ResNet34的皮肤病分类模型(Enhanced-ResNet34)。该模型在ResNet34的架构基础上，引入了自适应归一化层以提升小批量训练的稳定性，并构建了融合SE注意力机制与随机深度正则化的改进残差块，以增强关键病理特征的表达能力并降低过拟合风险。同时，在网络末端接入密集连接块以加强特征复用。实验采用HAM10000和ISIC2019两个公开数据集进行验证，通过迁移学习、高级数据增强及网络结构优化的逐步消融实验表明，本文模型在HAM10000验证集上的准确率达到92.6%，显著优于基线模型。此外，跨数据集泛化实验证明，该模型在ISIC2019数据集上仍保持87.0%的分类准确率，具备良好的鲁棒性与临床应用潜力。

Abstract: To address the issues of insufficient feature extraction and vanishing gradients in deep convolutional neural networks for dermoscopic image classification, this paper proposes an enhanced skin disease classification model based on ResNet34 (Enhanced-ResNet34). Building upon the standard ResNet34 architecture, the model introduces an adaptive normalization layer to improve training stability with small batches. Furthermore, an improved residual block integrating the Squeeze-and-Excitation (SE) attention mechanism and Stochastic Depth regularization is constructed to enhance the representation of critical pathological features and mitigate overfitting risks. Concurrently, a dense connection block is incorporated at the end of the network to strengthen feature reuse. The proposed model was evaluated on two public datasets: HAM10000 and ISIC2019. Through stepwise ablation experiments involving transfer learning, advanced data augmentation, and network structure optimization, the model achieved an accuracy of 92.6% on the HAM10000 validation set, significantly outperforming the baseline model. Additionally, cross-dataset generalization experiments demonstrated that the model maintained a classification accuracy of 87.0% on the ISIC2019 dataset, indicating strong robustness and potential for clinical applications.