A neural network based on knowledge distillation for classification of lung cancer with missing modalities

Lung cancer is a prevalent and life-threatening malignant tumor that significantly impacts human health. For clinical diagnosis of this cancer, positron emission tomography (PET) and computed tomography (CT) are frequently employed imaging methods. Using PET/CT images to accurate pathological classify lung cancer is crucial for effective clinical treatment. However, obtaining complete PET/CT data is often challenging due to clinical limitations. Therefore, developing a classification approach that addresses the problem of missing modalities has substantial clinical importance. This article introduces a novel neural network based on knowledge distillation to tackle the issue of missing modality, specifically designed for scenarios where only CT data is available. Our method effectively transfers knowledge from PET/CT data to CT data, capturing both intra-class and inter-class features. These features enable student models to learn robust feature representations from teacher models, using only CT data for inference. Our experiments resulted in an accuracy of 0.847. The findings suggest that our proposed method performs best when the inference data comprises only the CT modality.