A Novel Deep Learning Framework: Combining Two-Dimensional Principal Component Analysis and Convolutional Neural Network with Squeeze and Excitation Blocks for Skin Cancer Image Classification

This study proposes a skin image classification method by combining 2D-PCA for feature extraction with a Squeeze-and-Excitation (SE) enhanced Convolutional Neural Network (CNN) to improve performance on the HAM10000 and ISIC 2019 datasets. Initial experiments without regularization were conducted using learning rates of 0.01, 0.001, and 0.0001, epochs = 100, batch size = 32, and early stopping with a tolerance of 10 epochs. The best performance was obtained with a learning rate of 0.0001, achieving training accuracy of 99.92% and validation accuracy of 98.78% on HAM10000, with the optimal epoch at 48. Initial testing results reached 96.53% accuracy on HAM10000 and 95.43% on ISIC 2019, with corresponding high precision, Recall, F1-score, and Specificity. To enhance generalization, L2 regularization, dropout, and normalization were applied while keeping the same training settings. The best post-regularization results were obtained with a learning rate of 0.001 and dropout of 0.5, yielding training accuracy of 98.8% and validation accuracy of 97.55% on HAM10000. Testing after regularization showed improved performance: 96.80% accuracy on HAM10000 and 95.12% on ISIC 2019, with consistently higher precision, Recall, F1-score, and Specificity. Comparisons indicate that our method outperforms multiple state-of-the-art approaches, including EfficientNets, MobileNet, Modified Inception-v4, Decision Support System, Combined ResNet50-Exception, BILSK, 2D-PCA with various classifiers, and CNN-SE Blocks with SVMs, both before and after regularization. These results demonstrate that combining 2D-PCA with SE-CNN and proper regularization strategies effectively enhances classification accuracy, generalization, and stability in medical image analysis. This innovation enables efficient large-scale image analysis without sacrificing accuracy, providing practical value for early skin cancer detection, mitigating potential health risks, and supporting public health initiatives.