LPC-transformer: a training strategy optimized framework for multiclass pneumonia medical image classification

LPC-Transformer is a new framework designed for pneumonia classification on lung CT images, utilizing Swin Transformer as the backbone network. The framework was optimized through improved training strategies without requiring network deepening. The system uses a hybrid LMF loss combining Focal Loss and LDAMLoss to address class imbalance problems. The SCION optimizer is used to stabilize training and PolyReLU in MLP modules to model complex nonlinear features. When tested on the CL-COVIDset containing 11 classes of lung CT images, LPC-Transformer achieved significant improvements compared to the baseline SwinT model. Accuracy increased by 4.2%, recall by 4.4%, precision by 2.0%, and F1-score by 3.8%. The results validate the effectiveness of the proposed training strategy optimization.