Positron emission tomography - computed tomography (PET-CT) is now accepted as the best imaging technique to accurately stage lung cancer. The consistent and accurate interpretation of PET-CT images, however, is not a trivial task. We propose a discriminative, multi-level learning and inference method to automatically detect the pathological contexts in the thoracic PET-CT images, i.e. the primary tumor and its spatial relationships within the lung and mediastinum, and disease in regional lymph nodes. The detection results can also be used as features to retrieve similar images with previous diagnosis from an imaging database as a reference set to aid physicians in PET-CT scan interpretation. Our evaluation with clinical data from lung cancer patients suggests our approach is highly accurate.