Purpose: Regions of Cherenkov light attenuation, such as veins, have been observed in the imaging of breast cancer patients receiving external beam radiation therapy. This work aims to segment Cherenkov images by deep learning for the first time to identify in vivo fiducials, to explore biomarker-based real-time tracking by utilizing Cherenkov imaging in RT.
Methods: Cherenkov images of breast cancer patients receiving RT acquired in our IRB-approved clinical trial were used in this study. Regions of interest including the areola, tissue fold, veins and background were manually segmented as the ground truth. An Encode-Decoder network was designed as the backbone to process these images. In the Encoder module, the latent representation would be extracted from the input images and reserve the contextual information in the feature map. The Decoder module would reconstruct the images with label from the feature map. In order to improve the robustness in network, we consider removing the pure background pixels and utilizing random cropping and rotating to augment the training images in the preprocess stage.
Results: This architecture was pilot-tested on a limited data set of 11 patients, where regions of attenuation were clearly segmented in Cherenkov images. Quantitative evaluations of the accuracy and robustness will be presented with new patient data independent from what is used for training.
Conclusion: This study presents the first known work on deep learning applied to Cherenkov image processing. The efficient and robust veins segmentation enables real-time biomarker-based motion tracking in radiation therapy.