Medical image analysis suffers from the inherent difficulty in making reliable inference with limited contrast between structures such as tumors and their surrounding tissues. Relatedly, AI models are “black box” models with limited ability to provide explanation of inference, especially in terms of semantic segmentation and quantifying uncertainty of image registration for reliable estimation of deposited treatment doses. To address these two issues, we are developing new models, including distillation learning methods, to improve inference capability of AI models by learning to extract representations that better signal the contrast between foreground (e.g., tumor) and background (e.g., normal parenchyma). One example of these models called cross-modality educed distillation was successfully used for segmenting tumors, including those attached to the mediastinum, from CT and cone-beam CT images.