Path-CT Image Registration with Self-Supervised Vision Transformer for Lung Cancer

Computed Tomography (CT) imaging plays a crucial role in the diagnosis and treatment of cancer. Enhancing the interpretation of CT scans could greatly benefit radiologists by providing them with a definitive reference that highlights the tumor locations within the CT images. This has been achieved by registering pathology biomarkers for patients who have radical pre-operative CT scans. In this study, we assert that registering the cellular features in pathology images to CT images can produce images with more structural information about tumors and tissues than just identifying the tumor locations. This would significantly enhance the quality of subsequent analysis such as tumor segmentation, and reduce the variability in the interpretation of CT and pathology images. We initiate the process by training two self-supervised learning feature extractors for CT and pathology images, thereby enhancing the representation of both image types. Subsequently, we implemented a feature-matching network to achieve highly accurate registration between CT and pathology images. Our approach has yielded an accuracy rate of 72.5% on the lung dataset, surpassing the traditional method of accuracy of 65.9%.