Nowadays, lung cancer is the most common and lethal invasive tumor type in Chinese population, challenging overall health level. However, personalized early-stage treatment is currently still not widely implemented, and the choice of treatment highly depends on experience of physician. Based on deep learning and radiomics principles, deep-radiomics is important for establishing objective and promotable precision medicine plans. Among all aspects, the explainability of a model is critical for its usage in clinical practice. This paper discusses the technical aspects of explainable deep-radiomics in lung cancer, and analyzes challenges we are facing. Non-fully supervised learning methods, as a current hotspot in deep learning technology, can construct more trustworthy and practically valuable deep learning models through the co-design method of performance-interpretability. Medical artificial intelligence faces three core challenges in transitioning from the laboratory to hospitals: high-level cognitive demands, data privacy and generalization capabilities, and regulatory compliance. However, with appropriate design, non-fully supervised learning holds the greatest potential to bridge the gap between design and application, enabling broader adoption.