To effectively evaluate the radiation performance of RF front-end circuitries as demanded by the wireless industry, this article proposes a deep learning-based over-the-air (OTA) measurement system. By training a fully-connected deep neural network (FCDNN) with radiation measurements in some test points, we are able to accurately estimate the radiation performance of a RF circuitry in all 3D directions. To balance between the number of radiation measurements for FCDNN training and the estimation accuracy, we further propose to dynamically evaluate the accuracy of the trained model and increase the number of training radiation measurements, until the trained mode can satisfy a predefined accuracy. Experimental results show that the proposed OTA measurement system can accurately reconstruct the radiation performance of a RF circuitry with approximately 60% test points as compared to traditional methods. The proposed OTA measurement system can provide an accurate but cost-effective radiation measurement solution for the wireless industry.