LRCTNet: A lightweight rectal cancer T-staging network based on knowledge distillation via a pretrained swin transformer

Abstract

Rectal cancer, a prevalent malignant neoplasm within the digestive system, significantly jeopardizes patient health and quality of life. Accurate preoperative T-staging is critical for developing effective treatment strategies. In areas with limited medical resources, computed tomography (CT) has become the norm because of its popularity and economy and is an important method for the initial diagnosis of disease. Despite major advancements in computer vision in recent years, large-scale models have high demands on hardware and datasets, making them difficult to use and deploy in resource-limited environments. To address this challenge, we designed two lightweight modules, LightFire and ResLightFire, and developed a lightweight rectal cancer T-staging network (LRCTNet). On this basis, we leveraged the swin transformer, transfer learning and knowledge distillation techniques to optimize the classification performance of the LRCTNet. The experimental results revealed that LRCTNet achieved a classification accuracy of 95.79%, precision of 93.91%, recall of 93.48%, F1 score of 93.70%, and Matthews correlation coefficient (MCC) of 94.38% while containing only 0.407 million parameters, which were much higher than those of lightweight models such as SqueezeNet, MobileNet, and EfficientNet. These results indicate that the model achieves a low misclassification rate and a low rate of missed detections, ensuring balanced performance in classification. The lightweight design of LRCTNet enables efficient deployment in resource-constrained environments without sacrificing accuracy, making it a valuable tool for rectal cancer diagnosis.