LPC-transformer: a training strategy optimized framework for multiclass pneumonia medical image classification
TL;DR
LPC-Transformer, a framework optimized via improved training strategies with Swin Transformer as the backbone, proposes an LMF hybrid loss integrating Focal Loss and LDAMLoss to mitigate class imbalance to mitigate class imbalance.
Abstract
Lung imaging enables direct visualization of lesions and is vital for pneumonia diagnosis. However, current deep learning models show unsatisfactory classification accuracy and generalization for multi-class lung disease recognition, limiting their clinical decision-support value. To tackle this issue, we present LPC-Transformer, a framework optimized via improved training strategies with Swin Transformer as the backbone; performance gains are realized without deepening the network. We propose an LMF hybrid loss integrating Focal Loss and LDAMLoss to mitigate class imbalance. Furthermore, the SCION optimizer is adopted to stabilize training, and PolyReLU is used in MLP modules to model complex nonlinear features. Evaluated on CL-COVIDset containing 11 classes of lung CT images, our method is compared with mainstream benchmark networks. Experiments reveal that LPC-Transformer surpasses the baseline SwinT model substantially, with accuracy, recall, precision and F1-score improved by 4.2%, 4.4%, 2.0% and 3.8%, respectively. These results validate the effectiveness of the proposed training strategy optimization.