The tongue image is an important physical information of human, it is of great importance to the diagnosis and treatment in clinical medicine. Herbal prescriptions are simple, noninvasive and low side effects, and are widely applied in China. Researches on automatic construction technology of herbal prescription based on tongue image have great significance for deep learning to explore the relevance from tongue image to herbal prescription, and can be applied to healthcare services in mobile medical system. In order to adapt to the tongue image in a variety of photographing environments and construct the herbal prescriptions, a neural network framework for prescriptions construction is designed, which includes single / double convolution channels and fully connected layers, and propose the mechanism of auxiliary therapy topic loss to model the therapy of Chinese doctors then alleviate the interference of sparse output labels to the diversity of results. The experimental data include the patient tongue images and their corresponding prescriptions from real world outpatient clinic, and the experiment results can generate the prescriptions that are close to the real samples, which verifies the feasibility of the proposed method for automatic construction of herbal prescription from tongue image. Also, provides a reference for automatic herbal prescription construction from more physical information (or integrated body information).