Abstract
Lung cancer is a significant cause of cancer-related deaths globally. X-ray image has been widely used for first-stage screening as it is affordable and widely available. Recently, with the development of the gas sensor IC chip, low-cost enose sensing exhaled breath from patients can potentially be used for the first-stage screening in the near future. We propose a share-embedding model combining x-ray images and enose sensory signals to diagnose lung cancer. Our model contains two branches: the image branch and the enose branch. Since the lack of the enose data, we try to use the pretrained image model to guide the enose branch to align toward the embedding space that the image model learned. Our share-embedding model is designed to be robust to domain shifts across devices and environments. To further improve performance, we use semi-supervised learning with instance weighting to transfer the model to the unlabeled target domain. To train and evaluate the performance, we collect the first paired X-ray images and enose data across multiple devices and clinical environments. In the experiments, our method outperforms each individual branch and a feature concatenation fusion method. In the cross-device setting, our method leveraging semi-supervised learning achieves the best performance.