GNNs Python script for TCM fingerprints

We employed deep residual networks, specifically ResNet-18, ResNet-34, and ResNet-50 architectures, to analyze and categorize chromatographic fingerprints of traditional Chinese medicines (TCM). 1. Model Architecture: ResNet Variants: Implemented ResNet-18, ResNet-34, and ResNet-50 to manage different complexities in image data. Pre-trained Weights: Initialized with weights from the ImageNet dataset to improve generalization across medical images. Custom Top Layer: Included a 50% dropout rate and a fully connected output layer tailored for classifying chromatographic fingerprints. 2. Data Preprocessing and Augmentation: Standardization: Resized all images to 384x384 pixels and normalized using standard values. Augmentation: Applied random flips, rotations, color jittering, and affine transformations to increase model robustness. 3. Training Setup: Loss Function: Utilized cross-entropy loss for effective training. Optimizer: Used Adam optimizer with initial learning rate of 0.0001 and weight decay. Early Stopping: Implemented to halt training when no improvement in validation loss is observed, preventing overfitting. 4. Validation and Performance Metrics: Dataset Split: Data divided into training (70%), validation (15%), and testing (15%) sets. Metrics: Monitored accuracy, recall, F1 score, and precision to evaluate model performance. 5. Practical Application: Image Prediction Workflow: Developed a workflow for loading trained models, preprocessing input images, and classifying new images, facilitating practical use in diagnostics.本研究采用深度残差网络（deep residual networks），具体选用ResNet-18、ResNet-34及ResNet-50架构，对传统中药（traditional Chinese medicines, TCM）的色谱指纹图谱开展分析与分类工作。 1. 模型架构： ResNet变体：采用ResNet-18、ResNet-34与ResNet-50以适配图像数据的不同复杂度。 预训练权重：采用ImageNet数据集的预训练权重进行初始化，以提升模型在医学图像场景下的泛化能力。 自定义顶层网络层：设置50%的丢弃率（dropout），并构建适配色谱指纹图谱分类任务的全连接输出层。 2. 数据预处理与增强： 标准化处理：将所有图像统一调整至384×384像素，并采用标准值进行归一化操作。 数据增强：应用随机翻转、随机旋转、色彩抖动及仿射变换，以提升模型的鲁棒性。 3. 训练配置： 损失函数：采用交叉熵损失函数以保障训练的高效性。 优化器：选用Adam优化器，初始学习率设为0.0001，并配置权重衰减项。 早停机制：当验证集损失未出现改善时终止训练，以避免模型过拟合。 4. 验证与性能指标： 数据集划分：将全部数据划分为训练集（70%）、验证集（15%）与测试集（15%）。 评估指标：监控准确率（accuracy）、召回率（recall）、F1分数（F1 score）与精确率（precision），以全面评估模型性能。 5. 实际应用： 图像预测工作流：开发了一套完整的工作流程，涵盖加载训练完成的模型、预处理输入图像及对新样本图像进行分类，可便捷应用于实际诊断场景。